I can’t embed it here, but if you click this form, you’ll find 2–5 quick questions about which kinds of posts you value most and a few inexpensive course ideas I’m weighing. Images of the questions are below.

[Link to survey]

One more thing: I’ll be in China (the Shanghai, Hangzhou, and Wuxi area) from roughly mid-July through mid-August. If you are located anywhere in East Asia and you’d like a keynote, workshop, or just a conversation about AI, debate, or both, hit reply.

Stefan Bauschard
Stefan-Bauschard.com

Subscribe now

At this year’s US National Speech & Debate tournament, students participating in Congressional Debate will debate many bills and resolutions that are centered on underlying issues related to artificial intelligence.

These bills include items related to lethal autonomous weapons, facial recognition technology, generative AI in education, regulating the commercial use of generative AI, data center regulation, and strengthening domestic semiconductor supply chains.

In this post, I discuss the background of a bill — lethal autonomous weapons legislation — the pro and con arguments, and provide some advice for students who will be debating it. I hope it both gives you an overview of the complexity of the debate and also demonstrates how much “cognitive acceleration” (I made up this term) debate requires.

Student classroom debate over issues you are teaching is a great way to promote learning. Students need to research and prepare arguments about the issue you are teaching them. They have to articulate their arguments orally and defend them across a debate, eliminating the ability of AI to simply do their work. They also develop essential skills in argumentation, collaboration, and communication. They can also learn how hard it is to write legislation (terms, workability, etc). [The legislation students debate at tournaments is sample legislation written by the student participants.]

If you choose AI topics, this is a great way to promote student knowledge of AI without non-stop debates about what “AI literacy” and “AI fluency” mean, whether it should be taught in schools, who should teach it, how it should be taught, what it should displace, etc. The never-ending debates that aren’t doing anything for students. See GlobalAI Debates.

A Congressional debate-style format also allows a class of 12-24 students to participate.

Now, let’s analyze a bill the way a debater would.

Subscribe now

Part I — The Policy Debate

Background: how the U.S. arrived at this debate

The U.S. policy story is one of building autonomy faster while debating whether to restrain it. The governing document, Department of Defense Directive 3000.09, was the world’s first national policy on autonomous weapons when it was issued in 2012, and it was updated in January 2023.

It never banned anything — it created a review process and required “appropriate levels of human judgment over the use of force” (CRS). Michael Horowitz, a college National Debate Tournament (NDT) champion, who helped write the update, says in War on the Rocks that the 2023 revision was prompted partly by the Russia-Ukraine war demonstrating both the utility of AI-enabled weapons and their necessity, since electronic warfare jams remotely-piloted systems. Alongside the directive, the State Department issued a 2023 Political Declaration on Responsible Military Use of AI and Autonomy, and Congress layered on reporting requirements — the FY2025 NDAA already mandates an annual report on LAWS approval and deployment through December 2029 (CRS).

The money tells the direction of travel. The Brennan Center’s March 2026 explainer, “The Military’s Use of AI,” documents at least $75 billion in DoD AI spending since 2016 and $13.4 billion requested for autonomous systems in FY2026 alone, and the FY2027 budget reportedly proposes a steep further increase for autonomous warfare (Cipher Brief). In July 2025, the Pentagon signed deals with four frontier-AI companies — Anthropic, OpenAI, xAI, and Google — to develop military applications of their foundation models (Brennan Center). The U.S. is not drifting toward a ban; it is institutionalizing autonomy across its forces. That is the backdrop against which any prohibition would land.

The Pentagon-Anthropic dispute

The clearest sign that this debate has moved from the abstract to the concrete is the rupture between the Department of War and Anthropic, the maker of the Claude AI model. The dispute is worth understanding in detail because it is a live test of the exact question a ban turns on: who decides whether a fully autonomous weapon gets built and used.

According to the Brennan Center, Anthropic asked the military to promise it would not use Claude in weapons that identify and fire on targets without human input — fully autonomous weapons — or to conduct mass domestic surveillance of Americans by analyzing location records, financial information, and other large datasets. The request reportedly followed the military’s use of Claude in its January operation against Venezuela and the capture of Nicolás Maduro, and Claude had been integrated into the Maven Smart System used for targeting analysis (Brennan Center). The Pentagon refused the conditions and, in a statement from CEO Dario Amodei dated March 5, 2026, Anthropic confirmed it had been designated a “supply chain risk” to national security under the procurement statute it cites as 10 USC 3252. Anthropic is challenging the designation in court as not legally sound, argues the designation is narrow in scope, and frames its position as only two exceptions — fully autonomous weapons and mass domestic surveillance — while explicitly stating it does not believe a private company should be involved in military operational decision-making. Anthropic also said it would continue providing its models to the Department at nominal cost during any transition.

For this debate, the dispute matters in three ways.

First, it shows the human-judgment line is already contested at the contracting layer inside the United States, not only in Geneva — the fight is here and now.

Second, it puts a concrete definitional question on the table: Anthropic’s “fully autonomous weapons” exception is the same out-of-the-loop category any U.S. ban would target, so the dispute previews exactly the line-drawing problem a statute would face.

Third, it reframes the policy choice as one about who sets the limit — Congress by statute, the executive by directive, or vendors by contract. The Pentagon’s response suggests it views vendor-imposed restrictions on autonomous-weapons use as an intrusion on military prerogatives; Anthropic’s position is that it should not be compelled to enable either fully autonomous targeting or mass surveillance. Both framings are usable depending on which side of the ban you argue, and the episode is recent enough that you should pull the current status of the litigation before relying on it.

Subscribe now

What’s actually being debated

Most arguments about banning lethal autonomous weapons fail before they start because the two sides aren’t debating the same object. “Ban LAWS” can mean any of three very different things: a total prohibition on any weapon that selects and engages without a human; a narrow ban on systems that target people or operate fully out of the loop; or a ban only on offensive autonomy while preserving defensive systems. The strength of every pro and con below depends on which of these is on the table.

The technical baseline matters too. Autonomy runs on a spectrum the field describes as human-in-the-loop (a person authorizes each engagement), human-on-the-loop (the system engages under human supervision with an override), and human-out-of-the-loop (no human authorization, supervision, or intervention). The Congressional Research Service defines a lethal autonomous weapon as one that, once activated, “can select and engage targets without further intervention by a human operator” — the out-of-the-loop category — and states there is no agreed international definition. That absence is not a footnote; it is the central practical obstacle to any ban, and I’ll come back to it.

Current U.S. policy bans none of this. Department of Defense Directive 3000.09 governs autonomy through a review process and requires “appropriate levels of human judgment over the use of force” — not a human in the loop. Michael Horowitz, who helped rewrite the 2023 directive, argues in War on the Rocks that the phrase “human in the loop” appears nowhere in it, that no category of weapon is prohibited, and that R&D isn’t regulated at all. So the U.S. debate is about whether to adopt a ban, not whether to keep one.

Subscribe now

The global landscape

The international picture pulls two directions at once. Rhetorically, momentum runs toward prohibition: the UN General Assembly adopted Resolution 80/57 on December 1, 2025, reaffirming that any weapon that cannot be used in compliance with international humanitarian law must not be used, and Human Rights Watch reports more than 120 countries back treaty negotiations, with 96 states attending the first General Assembly meeting devoted to the issue in May 2025. The favored framework is the “two-tier” approach laid out by Austria’s Alexander Kmentt in Arms Control Today: prohibit systems that can’t comply with IHL, regulate the rest.

Operationally, the picture is the opposite. The Convention on Certain Conventional Weapons runs on consensus, and a handful of major powers — the U.S., Russia, India, and Israel — have used that rule to block a binding instrument for a decade (Kmentt). And the battlefield has lapped the diplomacy: the Cipher Brief’s reporting from Ukraine shows drones accounted for more than 80 percent of enemy targets destroyed by late 2025, with the defense ministry stating the goal is to remove human operators from the battlefield entirely. The norm is being written in Geneva and overwritten in Donbas at the same time.

Subscribe now

The case for a U.S. ban

The pro-ban case is strongest on accountability and weakest on enforceability. Run through the arguments in order of durability.

The first is the responsibility gap. International humanitarian law assumes a human agent who can be held responsible for a killing; a machine that selects and engages dissolves that agent. Jie Guo’s 2025 analysis in Ethics & Global Politics frames it through the philosopher Robert Sparrow: when an autonomous system commits what looks like a war crime, the programmer couldn’t foresee it, the operator didn’t control it, the commander cites technical complexity — so the violation is “procedurally unavoidable yet legally unpunishable.” Mary Ellen O’Connell’s 2023 essay for Ethics & International Affairs argues this makes the systems unlawful under existing law, ban or no ban.

The second is the black box. Because a learning system’s decisions can’t be predicted by its own designers, you cannot know at deployment whether it will comply with the law of war or the right to life (O’Connell). This isn’t a tuning problem; it’s structural, and it means a system certified as compliant on Monday may not be on Tuesday after it ingests new data.

The third is human dignity and meaningful human control — the frame the UN Secretary-General, the ICRC, and a large bloc of states have organized around. Even a perfectly accurate machine offends the principle that a person should decide to take a life — what O’Connell calls the problem of “mechanized killing,” and what the Vatican’s representative (cited in O’Connell) frames as decisions over life and death requiring compassion and insight no machine possesses. Berkeley’s Stuart Russell, who has made this case for a decade, draws the categorical line precisely: lethal autonomous systems “locate, select, and engage targets without human intervention,” unlike cruise missiles or remotely-piloted drones for which humans make all targeting decisions, which is why he calls them a third revolution in warfare after gunpowder and nuclear arms. The point isn’t that AI is bad; it’s that removing the human moral agent from the kill decision entirely — not merely distancing them, as a drone does — crosses a line earlier weapons didn’t. This is the argument that survives even if every technical objection is solved.

The fourth is that IHL’s core rules can’t be coded. Distinction and proportionality are interpretive judgments — weighing a tactical advantage against a destroyed school — not arithmetic. Guo (2025) reports that computer-vision systems for combatant identification achieve only 70–85 percent accuracy in cluttered environments, and the Brennan Center notes Project Maven’s algorithms correctly identified a tank about 60 percent of the time in good weather and 30 percent in snow (The Military’s Use of AI). Encoding proportionality, Guo argues, flattens moral complexity into programmable metrics.

The fifth is escalation. Machine-speed engagement compresses the decision loop below human reaction time and invites “flash war.” Guo (2025) cites wargaming showing autonomous interactions increase unintended-conflict initiation by 40–60 percent. The high-end version is nuclear: Lt. Gen. Jack Shanahan’s September 2025 Arms Control Today piece warns that AI bleeding into the sensors and decision-support systems around nuclear command — not on the launch button — creates automation bias and cascading errors that could distort a nuclear decision.

The sixth is proliferation. Autonomy is cheap and retrofittable. The Geneva Academy’s 2025 brief warns that many existing weapons can be retrofitted with autonomy and that low-end autonomous systems are far more attainable than sophisticated compliant ones, with some software already freely available. Resolution 80/57 names proliferation “to unauthorized recipients and non-State actors” as a core concern.

The seventh is bias and civilian risk, which doubles as a domestic-rights argument. Human Rights Watch’s April 2025 report “A Hazard to Human Rights“ argues these systems will migrate into policing and border control, where target profiles risk “digital dehumanization” and disproportionately harm communities of color. The Brennan Center adds that military AI used for surveillance can flag protected characteristics as security threats, raising First and Fourth Amendment problems when turned on Americans.

The eighth is automation bias and deskilling. Both Shanahan and Guo identify the tendency to over-trust machine output under time pressure — the dynamic behind the USS Vincennes shootdown of Iran Air 655 (Guo 2025) — which means even a human “in the loop” may rubber-stamp recommendations they can’t meaningfully evaluate. Over time, commanders lose the practical judgment the law assumes they have.

The ninth is the first-mover norm. Norms in weapons systems get set by the leading military power, so if the U.S. fields autonomous targeting, others follow — first peer competitors, then mid-sized powers, then non-state actors. The drone analogy is the precedent: the U.S. began weaponizing drones around 2001 in near-exclusivity, and by the 2020s New America’s database tracks more than ten countries that have conducted drone strikes and over three dozen with armed drones, while CNAS counts more than 30 nations with armed-drone programs and 90-plus with unarmed drones. The cat doesn’t go back in the bag. Advocates argue a unilateral ban creates the political space for treaty negotiation, while refusing to ban kills the treaty — and that the U.S., having helped draft the Geneva Conventions and the UN Charter, is the country whose restraint actually moves the norm.

Subscribe now

The case against a U.S. ban

The anti-ban case is strongest on reciprocity and feasibility and weakest on ethics. Same ordering.

The first is unilateral disarmament. A U.S. statute binds only the United States and is unverifiable against the actors the U.S. fears most. The CRS line is the realist core: the U.S. “may be compelled to develop the systems if U.S. competitors choose to do so.” The Lieber Institute (March 2026) judges a binding international instrument “slim to none” given great-power opposition — meaning a U.S. ban likely buys no reciprocal restraint at all. This also exposes a tension in the advocates’ first-mover frame: the proliferation that frame fears has arguably already begun without U.S. leadership — China’s pursuit of autonomy under its “intelligentized warfare” doctrine, Russia’s loitering munitions in Ukraine, and UN monitors’ account of a Turkish Kargu-2 possibly making the first autonomous engagement of human targets in Libya (cited in Guo 2025). If the norm is already breaking without the U.S. setting it, then unilateral abstention forfeits the capability without buying the norm-setting benefit — advocates can’t simultaneously claim the U.S. sets the norm and that adversaries are already racing ahead of it.

The second is military necessity, and it’s empirical now. The reason autonomy matters is that electronic warfare jams the human link: Horowitz notes the 2023 directive update was driven partly by Ukraine demonstrating that remotely-piloted systems get cut off, so autonomy becomes the only way to operate in a communications-denied environment. The CRS adds that LAWS are valued precisely for “communications-degraded or -denied environments.” In an Indo-Pacific fight against a peer adversary, a human-in-the-loop mandate can mean ineffective forces.

The third is that the technology is already decisive, not speculative. The Cipher Brief’s Ukraine reporting — 80 percent of kills by drone, units at 30–60 percent strength, robot positions held for 45 straight days — frames autonomy as a manpower necessity, and even argues that insisting on human-in-the-loop can be less ethical when it means slower casualty evacuation and more dead operators.

The fourth is that a ban sweeps in defensive systems the U.S. already depends on. Per Horowitz, the Phalanx Close-In Weapon System has been deployed since 1980, switches to an automatic mode that engages incoming threats faster than a human can, and was used in the Red Sea against Houthi missiles. Scientific American lists Iron Dome, Phalanx, and the German NBS Mantis as already-deployed defensive autonomy, and Paul Scharre has noted at least 30 countries field automated defensive systems including Aegis and Patriot. A clean ban on “select and engage without human intervention” pulls all of them in. There’s a deeper irony here that opponents can press: the moral premise of the ban is that the wrong lies in the absent human — but point defense against a supersonic missile is precisely the case where human reaction is physically impossible and the moral stakes of automation are lowest. The principle that drives the ban argues hardest for exempting exactly the systems a blanket ban catches, so the principle and a total prohibition point in opposite directions.

The fifth is speed against saturation. Defensive autonomy exists because incoming missile and drone salvos arrive faster than a human can authorize each intercept; Ukraine’s Octopus interceptor (Cipher Brief) destroys incoming drones without per-shot human approval. Requiring human approval of every intercept in a saturation attack, the argument runs, gets more people killed, not fewer.

The sixth is the precision claim. The U.S. government’s position, summarized in the CRS primer, is that automated targeting “can allow weapons to strike military objectives more accurately and with less risk of collateral damage.” If that’s even partly true, a blanket ban could increase civilian harm — turning the humanitarian argument against the ban.

The seventh is the capability gap and deterrence. Russia and China are investing heavily, and the FY2027 budget reportedly proposes $54.6 billion for autonomous warfare and a jump for the Defense Autonomous Warfare Group from roughly $225 million to tens of billions (Cipher Brief). Freezing U.S. development while adversaries accelerate, opponents argue, weakens deterrence rather than strengthening norms.

The eighth is industrial and economic. The Brennan Center documents that Palantir and Anduril recorded their largest-ever defense revenue in 2025 — $903 million and $912 million respectively — and that the Pentagon has signed foundation-model deals with Anthropic, OpenAI, xAI, and Google. A ban disrupts a defense-industrial base now organized around autonomy, with the Pentagon-Anthropic blacklist dispute showing how contested the terrain already is.

The ninth is that the definition makes a ban unworkable — which is really a practicality argument, so it gets its own section.

Subscribe now

The practicality question

This is where the debate is actually decided, and where both the bumper-sticker pro and the bumper-sticker con tend to collapse. Five problems, in order of severity.

The first is definitional. “Without human intervention” sounds crisp but isn’t. The CRS says there is no agreed definition; Horowitz argues DoD deliberately abandoned “in/on/out of the loop” language because it falsely implies continuous tactical oversight that even conventional precision weapons lack. A fire-and-forget missile already engages a human-designated target autonomously. Where, exactly, is the line between that and a banned system? A ban built on an undefined phrase invites years of definitional litigation and gives adversaries — and contractors — room to reclassify around it.

The second is the offense/defense line, which is the definitional problem in its most acute form. IEEE Spectrum argues that “offensive,” “autonomous weapon,” and “meaningful human control” all lack common definitions, that it’s hard to separate offensive from defensive weapons, and that autonomy has been used in offensive weapons for decades. Any ban that wants to preserve Phalanx and Aegis has to draw the offense/defense line the entire literature says can’t be drawn cleanly — and the same target-recognition autonomy powers an offensive strike drone and a defensive interceptor alike.

The third is verification and enforcement. A domestic U.S. statute is enforceable against the United States but invisible against adversaries, and the CCW’s decade of consensus-blocked failure (Kmentt) shows how hard multilateral verification is. Autonomy has no observable signature the way a missile silo does — the difference between a compliant and a prohibited system is often a software setting. So a ban that isn’t multilateral and verifiable is a ban that constrains only the side willing to be constrained.

The fourth is “control in name only.” Even a ban that mandates a human in the loop may not deliver real control. The Cipher Brief’s central finding is that human oversight can be “preserved in name” while the conditions that make it meaningful — time, attention, comprehensible volume — have eroded: Ukraine’s Avengers platform surfaces up to 12,000 targets a week, and exhausted analysts “rubber-stamp” recommendations they can’t re-evaluate. A statute that requires a human signature without addressing cognitive overload buys the appearance of accountability, not the substance — which means a poorly designed ban can fail on its own terms.

The fifth is dual-use and retrofit. The Geneva Academy brief notes that autonomy is largely a software layer on top of existing hardware, with some of the underlying software freely available. You cannot ban a capability the way you ban a chemical agent when the capability is code that can be added to a commercial drone. This is also why the funding logic in most ban proposals — redirect “offensive autonomy” money to “defensive AI” — tends to be incoherent: the two run on the same stack.

Subscribe now

The bill on the table

The proposal in front of the chamber converts the general debate into four concrete statutory choices, and it inherits the practicality problems above in specific form. The bill bans U.S. military and federal law-enforcement research, development, procurement, and deployment of weapons that select and engage targets without human intervention; reallocates existing LAWS funding to defensive AI and cybersecurity; requires the Department of War to report annually to the House Armed Services Committee; and takes effect January 1, 2027.

Two things the bill gets right. It anchors on the out-of-the-loop definition, which tracks the CRS framing rather than reaching for everything with an algorithm in it. And its reach into federal law enforcement is responsive to a genuine regulatory gap that the battlefield-only framing misses.

Four things break on contact with the analysis above. The first is scope: the bill bans “deployment,” not just future development, and “without human intervention” is undefined — so on the effective date it arguably sweeps in legacy defensive systems like Phalanx, Aegis, and Patriot, which engage incoming threats faster than a human can authorize (Horowitz, War on the Rocks). Carving them out reopens the offense/defense line that IEEE Spectrum says can’t be drawn cleanly. The second is the funding mechanism: redirecting “existing LAWS funding” to “defensive AI” assumes a clean split between offensive and defensive autonomy, when they run on the same stack — and if current policy classifies no fielded system as a prohibited LAWS, “existing LAWS funding” may be a null set that funds nothing. The third is enforcement: the bill pairs a ban with an annual report but specifies no penalty, no inspector, and no remedy when a system crosses the line, and the reporting partly duplicates the existing FY2025 NDAA mandate for an annual LAWS report through December 2029 (CRS). The fourth is timing and reciprocity: a unilateral ban effective January 2027 freezes U.S. development precisely as competitors accelerate, against a budget the Cipher Brief puts in the tens of billions for FY2027.

The bill, in other words, is the blunt version of a defensible idea. It legislates the principle that machines shouldn’t make the kill decision, but it does so with a definition too broad to administer, a funding theory that contradicts itself, and no teeth. The narrower instrument described in the next section would deliver most of the bill’s moral payload without most of its practical failure modes. (For a round-by-round breakdown of the bill — advocate and opponent cases, cross-examination, and a strategic verdict — see Part II below.)

Subscribe now

Where the serious middle lands

Strip the slogans and the credible positions converge on something narrower than a total ban and firmer than the status quo. The ethical case against a machine making the kill decision on a human is close to a moral consensus — even DoD reserves an explicit human-in-the-loop requirement for nuclear weapons (Horowitz), which concedes the principle at the highest-stakes level. But a total, unilateral ban founders on the definitional problem, the defensive-systems sweep, the verification gap, and the reciprocity problem.

The defensible policy, which Guo, Kmentt, and even Horowitz approach from their different directions, has four parts: prohibit autonomous targeting of humans and fully out-of-the-loop offensive systems; preserve human-supervised defensive autonomy; require genuine human judgment with explicit safeguards against cognitive overload rather than a nominal signature; and put the diplomatic weight behind the two-tier international instrument rather than going it alone, since a ban that doesn’t bind adversaries mostly disarms the United States. The hard truth underneath the whole debate is that this is a values choice dressed as a technical one: how much capability is the U.S. willing to forgo, against accelerating competitors, in exchange for the ethical and normative gains of leading. Pretending the practicality problems away — in either direction — is the one move that guarantees a bad answer.

Subscribe now

Part II — Congressional Debate Round Prep

This part uses Congressional Debate conventions: “advocates” argue the bill should pass, “opponents” argue against, and the voice is tactical — written for a competitor who may be handed either side. It re-approaches the bill from Part I as case construction, cross-examination, drafting analysis, and a strategic verdict.

1. A Bill to Prohibit Lethal Autonomous Weapon Systems and Reallocate LAWS Funding

What the bill does

The bill bans all U.S. military and federal law enforcement research, development, procurement, and deployment of weapons systems that select and engage targets without human intervention. It reallocates existing LAWS funding to defensive AI and cybersecurity, requires the Department of War to report annually to the House Armed Services Committee, and takes effect January 1, 2027. The factual baseline both sides work from: current U.S. policy does not ban these systems — Department of Defense Directive 3000.09 governs them through a review process and requires “appropriate levels of human judgment over the use of force,” and the Congressional Research Service states plainly that “U.S. policy does not prohibit the development or employment of LAWS.” This bill would be a reversal of the status quo, not a codification of it.

Subscribe now

The strongest case for the bill

Advocates win on the moral architecture of accountable killing — the ground where the bill is most defensible and the chamber is most receptive.

The first argument is the responsibility gap. International humanitarian law assumes a human agent who can be held responsible for a killing, and a machine that selects and engages targets dissolves that agent. Guo’s 2025 analysis in Ethics & Global Politics frames it as a “responsibility gap“ — when an autonomous system commits what looks like a war crime, the programmer couldn’t foresee it, the operator didn’t control it, and the commander cites technical complexity, so no one is culpable. If you’re advocating, this is your lead, because it reframes the bill as preserving law rather than restricting capability.

The second argument is the black box. O’Connell’s 2023 piece for Ethics & International Affairs, “Banning Autonomous Weapons,” argues that because a learning system’s decisions can’t be predicted by its own designers, you cannot know at the moment of deployment whether it will comply with the right to life. Pair it with the documented failure: a UN panel found Turkish Kargu-2 drones attacked targets in Libya based on anomalous signatures, the case Guo (2025) cites as the responsibility gap confirmed in the field.

The third argument is escalation. Machine-speed engagement compresses the decision loop below human reaction time and invites “flash war.” The high-end version is nuclear: Lt. Gen. Shanahan’s September 2025 Arms Control Today piece warns that AI bleeding into the sensor and decision-support systems around nuclear command creates automation bias and cascading errors even if no one puts AI on the launch button. The November 2024 Biden-Xi agreement that AI must never supplant human judgment in nuclear launch decisions (also in Shanahan) is your evidence that even the U.S. concedes the principle at the top of the ladder.

The fourth argument is the law-enforcement clause, which most of the chamber will skip. The bill reaches federal police use, and Human Rights Watch’s April 2025 report “A Hazard to Human Rights“ argues these systems will migrate into policing, where the right to life and non-discrimination protections are stronger than IHL and biased target profiles risk “digital dehumanization.” If you’re advocating and the room saturates on the battlefield framing, pivot here — it’s fresh ground.

The fifth argument is leadership and momentum. The U.S. is in a shrinking minority. The UN General Assembly adopted A/RES/80/57 on December 1, 2025, reaffirming that any weapon that cannot be used in compliance with IHL must not be used, and Human Rights Watch reports more than 120 countries back treaty negotiations. Going first reclaims the norm the U.S. helped build.

Subscribe now

The strongest case against the bill

Opponents win on reciprocity and feasibility — and the bill hands them a clean procedural objection most advocates won’t see coming.

The first argument is unilateral disarmament. A U.S. statute binds only the United States and is unverifiable against the actors the U.S. fears. The CRS primer’s own line — the U.S. “may be compelled to develop the systems if U.S. competitors choose to do so” — is the realist core, and the Lieber Institute (March 2026) judges a binding international instrument “slim to none” given great-power opposition. If you’re opposing, this is your spine: the bill costs the U.S. the capability and buys no reciprocal restraint.

The second argument is military necessity, and it’s empirical now, not hypothetical. The Cipher Brief’s reporting from Ukraine’s frontline shows drones accounted for more than 80 percent of enemy targets destroyed by late 2025, with units operating at 30–60 percent strength and the defense ministry stating the goal is to remove human operators from the battlefield entirely. The reason autonomy matters: electronic warfare jams the human link, the point Michael Horowitz — who helped rewrite Directive 3000.09 — makes in his May 2025 War on the Rocks piece. In a communications-denied fight, a human-in-the-loop mandate can mean ineffective forces.

The third argument is that the bill bans defensive systems the U.S. already relies on. Per Horowitz, the Phalanx Close-In Weapon System has been deployed since 1980 and switches to an automatic mode that engages incoming threats faster than a human can — and it was used in the Red Sea against Houthi missiles. Scientific American lists Iron Dome, Phalanx, and the German NBS Mantis as already-deployed defensive autonomy, and former DoD official Paul Scharre has noted at least 30 countries field automated defensive systems including Aegis and Patriot. A clean ban on “select and engage without human intervention” sweeps all of them in.

The fourth argument is the procedural objection — run this early. The bill assigns no enforcement mechanism. The Department of War reports annually to House Armed Services, but reporting is not enforcement; there is no penalty, no inspector, no remedy when a system crosses the line. Worse, much of the architecture the bill claims to create already exists: the FY2025 NDAA already requires an annual report on LAWS approval and deployment through December 2029 (CRS). If you’re opposing, point out that the bill’s signature accountability feature is partly redundant and entirely toothless.

The fifth argument is the funding incoherence, which feeds the logical-flaws section below. The bill reallocates “existing LAWS funding” to “defensive AI,” but the Brennan Center’s March 2026 explainer, “The Military’s Use of AI,” documents $75 billion in DoD AI spending since 2016 and $13.4 billion requested for autonomous systems in FY2026 — and the offensive and defensive applications run on the same underlying autonomy. You can’t cleanly defund one and fund the other.

Subscribe now

Cross-examination questions

Questions for advocates to ask opponents:

1. “You say a ban is unilateral disarmament — but the U.S. already restricts itself on chemical and biological weapons that adversaries might cheat on. Why is autonomy the exception?”

2. “If Phalanx is purely defensive, the bill can carve it out — so isn’t your strongest objection just a drafting fix, not a reason to kill the bill?”

3. “You cite Ukraine’s 80 percent drone-kill figure. How many of those engagements were fully out-of-the-loop versus a human authorizing the strike?”

4. “When an autonomous system kills the wrong people, who do you court-martial — the coder, the operator, or the commander?”

5. “The November 2024 Biden-Xi agreement says AI must never authorize a nuclear launch. If the principle holds for nuclear weapons, why not for the lethal decision generally?”

6. “More than 120 countries back a treaty. If the U.S. is right and they’re wrong, why is the U.S. in the minority?”

7. “You want to keep building these. Name the adversary capability that a U.S. autonomous offensive weapon deters that a human-supervised one doesn’t.”

8. “You call this unilateral disarmament — but the bill bans development, not just deployment. Are you arguing the U.S. should spend money building weapons it never intends to field?”
   

   Subscribe now

Questions for opponents to ask advocates:

1. “Define ‘without human intervention.’ Does a fire-and-forget missile that homes on a human-designated target violate your bill?”

2. “Your bill bans ‘deployment.’ Does that pull Phalanx, Aegis, and Patriot off Navy ships on January 1, 2027?”

3. “You reallocate LAWS money to ‘defensive AI.’ The same autonomy powers both — how does the Department of War decide which dollar is offensive and which is defensive?”

4. “The FY2025 NDAA already mandates an annual LAWS report through 2029. What does your reporting requirement add?”

5. “If China and Russia keep building and the U.S. stops, what is your plan for the capability gap in a communications-denied fight?”

6. “Your bill has a ban but no penalty. What happens to a commander who deploys a prohibited system anyway?”

7. “You say human control is a moral imperative. The Cipher Brief shows analysts ‘rubber-stamping’ 12,000 machine-flagged targets a week. Is a human who can’t meaningfully review the targets actually control, or theater?”

8. “Your accountability argument says someone must answer for a wrongful killing. Commanders are already liable for the systems they deploy under the UCMJ and the law of war. Name what your bill adds to that chain that existing law doesn’t already cover.”

Drafting and definitional traps

The fatal flaw is “without human intervention.” The CRS primer says there is no agreed definition of LAWS internationally, and Horowitz argues DoD deliberately dropped “in/on/out of the loop” language because it falsely implies continuous tactical oversight that even conventional precision weapons don’t have. A statute built on the undefined phrase invites years of litigation over what counts.

The “deployment” verb is the second trap. Banning deployment, not just future development, is what sweeps in legacy defensive systems. If you’re opposing, read the verb out loud and ask whether the Navy strips Phalanx on the effective date. If you’re advocating, you must pre-empt this with a defensive carve-out for systems that exclusively engage incoming munitions — at which point you’ve reopened the offense/defense line that IEEE Spectrum says no one can cleanly draw.

The “existing LAWS funding” term has no defined baseline. If current policy doesn’t classify any fielded system as a prohibited LAWS — which is the CRS position — then “existing LAWS funding” may be a null set, and the reallocation funds nothing.

A smaller catch rewards the closest reader: the bill names the “Department of War” as the enforcing agency in Section 3 but reverts to “Department of Defense” in Section 3(B). Whatever your view of the 2025 rebrand, using two names for one agency in a single bill is a drafting inconsistency — flagging it signals you read the text more carefully than the room did.

Subscribe now

Logical flaws

The bill’s central contradiction is that it presumes a clean line between offensive autonomy (banned) and defensive AI (funded) while its own funding mechanism depends on that line being un-drawable. The same target-recognition autonomy that powers an out-of-the-loop strike drone powers an out-of-the-loop missile-defense interceptor — Ukraine’s Octopus interceptor in the Cipher Brief is exactly this. The bill bans the capability and then redirects money to the capability under a different name. Opponents should name this as self-defeating: the standard that disqualifies the weapon also disqualifies what the bill funds.

The second flaw is a non-sequitur in the advocates’ own framing. The case for the bill rests on preserving “meaningful human control,” but the empirical record the advocates rely on — the Ukraine data, the “rubber-stamping” of machine recommendations in the Cipher Brief — shows human control already eroding under target volume regardless of any legal requirement. Mandating a human in the loop without addressing cognitive overload produces nominal control, not real control, which means the bill may not actually deliver the accountability it promises. This cuts against advocates if opponents are sharp enough to turn the evidence around.

The third flaw inverts the bill’s own moral premise. The case for the bill says the wrong lies in the absent human — but point defense against a supersonic missile or a saturation drone swarm is precisely the case where no human can react in time, so it is the least morally fraught use of autonomy and the one the bill’s own principle should want to protect. The bill bans it anyway, alongside offensive targeting. If you’re opposing, this is the cleanest way to show the text and the theory point in opposite directions: the systems most defensible on the advocates’ own logic are the ones the bill sweeps in.

The fourth flaw is a non-sequitur in the first-mover argument. Advocates claim that if the U.S. fields LAWS everyone follows, so the U.S. must abstain to create treaty space. But the opponents’ own evidence — China’s “intelligentized warfare,” Russian loitering munitions in Ukraine, and the UN account of a Kargu-2 engagement in Libya (Guo 2025) — says proliferation has already begun without U.S. leadership. If the norm is already breaking, “U.S. restraint sets the norm” collapses: the premise that the U.S. is the first mover is false, so unilateral abstention forfeits capability without buying the norm-setting benefit. Advocates can’t hold both “the U.S. sets the norm” and “adversaries are already racing ahead.”

The fifth flaw is the currency trap. Every load-bearing number in this debate is moving. The FY2027 budget reportedly proposes $54.6 billion for autonomous warfare and a jump for the Defense Autonomous Warfare Group from roughly $225 million to tens of billions (Cipher Brief). The Pentagon-Anthropic blacklist dispute (Brennan Center) is live. Pull the current figures the week before you speak; a stale number is a CX liability.

Subscribe now

Verdict / how to play it

The chamber will saturate on the advocate side — the moral case against killer robots is intuitive and the evidence base is rich, so most speeches will be some version of “machines shouldn’t decide who dies.” That makes the competent opponent speech the rarer and higher-scoring one, and the bill hands it three real points: the definitional overbreadth, the defensive-systems sweep, and the no-enforcement procedural gap. I’d open opposition with the procedural objection — it’s clean, factual, and most of the room missed that the bill has a ban with no penalty and a reporting requirement that partly already exists.

On the advocate side, the single highest-leverage move is to narrow the bill in your own framing before anyone attacks it: concede the defensive carve-out, anchor on out-of-the-loop targeting of humans, and lead with the responsibility gap (Guo 2025) rather than a blanket ban. The advocate who debates the bill as written loses to the Phalanx question; the advocate who debates the principle wins the room.

The cross-apply: if this docket contains any AI-governance or military-spending bill, the offense/defense-line problem and the “authorization or reallocation without a workable definition” move travel directly. Run the same enforcement-agency and definitions checks on those.

Subscribe now

US Policy

Defense Primer: U.S. Policy on Lethal Autonomous Weapon Systems

Legal Accountability for AI-Driven Autonomous Weapons

The Military’s Use of AI, Explained

International Action

Geopolitics and the Regulation of Autonomous Weapons Systems

156 states support UNGA resolution on autonomous weapons

General

AI Arms Race: How Autonomous Systems Are Reshaping Deterrence and Escalation Dynamics

International

LAWS Resolution of the UN General Assembly

UN: Start Talks on Treaty to Ban ‘Killer Robots’

The Pentagon’s $54 Billion Bet on Autonomous Warfare

Lethal Autonomous Weapons Bad

A Hazard to Human Rights: Autonomous Weapons Systems and Digital Decision-Making

The ethical legitimacy of autonomous Weapons systems: reconfiguring war accountability in the age of artificial Intelligence
Recent advancements in artificial intelligence have intensified debates on deploying Autonomous Weapons Systems (AWS) in warfare. Proponents justify AWS on grounds of(1) enhanced military efficiency and reduced soldier casualties, (2) improved compliance with international humanitarian law (IHL) through algorithmic precision, and (3) operational necessity in high-threat environments. This paper critically examines these arguments, contending that they fail to establish the ethical legitimacy of AWS. It argues that AWS fundamentally undermine moral accountability in war, exacerbate risks to civilians, and corrode human agency in lethal decision-making. The analysis concludes that existing ethical and legal frameworks cannot adequately govern AWS, necessitating a reconfiguration of accountability paradigm.

LAWs Weapons Ban Bad

Banning Autonomous Weapons: A Legal and Ethical Mandate

Autonomous Weapon Systems: No Human-in-the-Loop Required, and Other Myths Dispelled

Human vs. Machine: Operational Realities from Ukraine’s Frontline

Nuclear Arms & AI

Artificial Intelligence and Nuclear Command and Control: It’s Even More Complicated Than You Think

Subscribe now

Not everyone yet realizes just HOW very important AI is going to be as we move through AGI and superintelligence, but everyone and their uncle… and their religious leader … is recognizing AI is a super huge deal and the theme of the times.
— Dr. Ben Goertzel

In this week’s update, we focus on the societal-level changes society is facing as a result of AI.

We begin with an overview of Demis Hassabis’ AGI claims, with a focus on his claims about next-generation agents and that AGI could arrive as early as 2029. This is the same year Ray Kurzweil projected back in the 1990s.

We also discuss rapid advances in mathematics, applications in the life sciences, and what it means for AI to be a “utility.”

We then spend most of the podcast discussing the Pope’s AI Encyclical, with a focus on its call for educators to lead in this era.

All the links and additional updates are available for paid subscribers.

Subscribe now

Read more

Subscribe now

__

Chapter 1: Welcome and Introduction to Rich Brown

Anand: Well, we’re excited to welcome you back and to have this conversation with Rich Brown. Rich is a Marine veteran and small business owner here in Fredericksburg, Virginia. He’s also the founder and president of the nonprofit Guideon Foundation, which supports veterans and first responders by sponsoring events and training for those groups and young people in the community. I met Rich recently after a speaking event and had a chance to meet and talk with him about his work. Welcome to the podcast, Rich. We appreciate you being here.

Rich: Thanks, Don. I appreciate you giving me your time.

Anand: Awesome. So we first connected over a discussion about agentic AI, and you were the one in the room who had a much better story about using agentic AI and a better experience than anybody else. Tell us a little bit about how you’ve been using AI, agentic AI, in your work over the last couple of months.

Rich: Sure. When we met up in person, I told you that I was a little ashamed that I had talked about my spending habit with agentic AI as if I’d worn it almost as a badge of honor. All in total over the last two months, I’ve spent somewhere between two and four thousand dollars in Anthropic tokens. But really, I shouldn’t celebrate the fact that I don’t budget well for agentic AI.

I have gotten much better. I’ve created an infrastructure in which my agents are a lot smarter about the tokens we’re using. And I think I said this when we met — I use this metaphor to tell somebody else about the setup, and the fact that if you’re speaking just about Anthropic, obviously there are other services, but if you’re using Haiku, Haiku is about on the level of a college graduate in terms of what it’s able to produce. If you’re using Sonnet, that’s more like somebody who has graduated from college and has some experience actually doing the work you’re asking it to do.

Then Opus 4.7 is like asking — it’s like paying someone who’s famously good at the thing you’re asking it to do. And I’ve been very, very poor about my spending when it comes to having someone famously good manage my calendar and my day-to-day functions. My famously good writer doesn’t need to be calculating my calories for the stuff I’m eating on a daily basis. So knowing which model you’re using and being deliberate about the usage is smart — and smarter than I have been, unfortunately.

Anand: Well, now we know why they had to move to the Colossus data center — because of all the demand in Fredericksburg coming from your agentic workflow. So tell us how you’ve been using it in business, because this is something I think a lot of educators don’t really have a good appreciation for — how transformative it’s been for some workflows.

Rich: Sure. To back up a little bit, I own two companies and I serve as president of a nonprofit. So we could take any path you want, whether it’s the executive protection business or the gym or the nonprofit. I can say for the nonprofit, right off the bat — because it’s something I’ve been working on lately — we have a Service Leadership Academy that we’re hosting this summer where we’re going to take 35 rising high school seniors and teach them leadership from the perspective of veterans and first responders in the community, and some guest speakers from around the country.

I like to move fast. Whereas a decade ago it would have cost dozens of thousands of dollars to put together the curriculum and have it all vetted and reviewed and so on, I’m speaking with an agentic AI agent and developing the curriculum rapidly. I’m using source material I take a lot of confidence in — whether it’s Jordan Peterson, Jocko Willink, books like Tribal Leadership, or the various team-leading books from military veterans — using that as the foundation.

Where I think a lot of people screw up with AI is they’ll make a good prompt — they’ll tell the AI the role they want it to fill and then they give it a well-crafted prompt, and then they ask it to produce something like “develop the curriculum for this eight-week course.” That’s good. That’s above-average usage. That’s not just using AI as an advanced Google search. But where they don’t go the next step is to say — for example, this curriculum I’m developing, we’re going back and forth on about 400 questions that my agent is asking me about the curriculum before we ever write a line of text. Really fleshing out your ideas and the intent and the application — I think that’s where a lot of people fall short. It’s one thing to use AI to answer questions or to solve homework problems. It’s another thing to help it really dive deep into what it is you want it to produce.

Subscribe now

Chapter 2: The Cost of AI Experimentation and Smarter Token Use

Anand: And I was going to say — I think it’s definitely true in my experience that you should also get some free Claude tokens for your claim that it’s as good as a professional. I think I’m an unpaid fanboy spokesman.

Stefan: There’s got to be a way to maybe help monetize that with some short clips. But I think that’s definitely the case. And how are you using it? That curriculum example is something I think a lot of our listeners can relate to. But I’ve also heard — and this is incidental — I’ve heard from a number of kids who graduated last year or so, and they said, “Oh, our professors told us we can’t use AI. Don’t use AI in school, because you’re never going to get a job if you use AI at work, or if you tell the people who are hiring you that you used AI.” And obviously that’s kind of not the case. What are some ways that you or your employees you’re training are using AI, or that you’re encouraging them to use it?

Rich: Well, first I want to address the elephant in the room. I don’t want this to sound like a stab at you directly, because you work within the academic field — but I think a lot of professors are telling their students not to use AI when they’re developing anything for their academic pursuits. And unfortunately, it’s probably going to be the teachers and the professors who, in 10 to 20 years, aren’t going to have a job, all due to AI. I think the schools are going to have to figure out real quick what the future of academics looks like, because they’re not keeping up with what’s out there and what’s possible. I admire you two for trying to get ahead of that curve, and I think that’s what you’re doing with this podcast.

Your question was how I’m encouraging my employees to use AI. That’s a good question. I own an executive protection business — which is a really almost pretentious way of saying bodyguarding. But it is the very professional bodyguard. We’re not hoodlums protecting rappers — we’re protecting Fortune 50 CEOs. We’ve started using AI a lot in doing what’s called advance work.

Advancing any objective is when — let’s say I’m protecting a CEO who’s going to be meeting with the president at the White House. They’re going to fly into Dulles from probably California, and we’re going to pick them up at Dulles, drive them to their hotel, take them to whatever meetings they have scheduled in the area, meet with the president, maybe spend another day in town, and then get back on an airplane and head back to California. We will spend up to a week or two going to all of the places they plan on visiting before their actual visit. We’re looking for ingress and egress routes, all the different safety concerns, how our principal may be vulnerable in each venue, developing relationships with local security and local law enforcement — so that we are 100% ready long before they ever set foot in that room.

Rich: A decade ago, you had to physically go to all of those places, shake all those hands, get the phone numbers for all of those people. We’re able to do it much faster and easier and do most of our homework first. I’ve got an agent whose entire role is — I give them a location, and they immediately produce, if the information is public, the director of security for that location, and the nearest police, law enforcement, fire, and the closest emergency venue — all of that, if I just type in a location, because it knows what I’m looking for. Does that kind of answer your question?

Stefan: Yeah. One thing I’ve found doing this for a while is that just talking to people, you learn about all different types of use cases. The first time, I think it was in the class I was teaching at Mary Washington, a student said, “Oh, I have trouble communicating with my roommate, so I use it to help me draft my text.” I’m like — people are using this for so many different things. It’s really going to explode. And that’s another unique use case probably many people didn’t know about till right now.

Subscribe now

Chapter 3: Using AI to Build Curriculum for a Service Leadership Academy

Rich: I’m not ashamed to tell you — and maybe I shouldn’t say this publicly — but I have an agent that acts as sort of an intermediary between me and my ex-wife in terms of custody matters and that sort of thing. It typically speaks from the point of view of an attorney and that kind of mindset. What I’ve done is created an agent where, if I give it a screenshot of a conversation, it has all of the history of the last two or three years’ worth of text-message exchanges between me and my ex-wife. If she sends me a message, I send a screenshot to my agent, and it gives me a kind of coaching of how I should respond.

What it’s looking for is: how do I stay — I’m not creating conflict, I’m trying to be as amiable as possible and keep 100% focus on what’s best for my daughter. It’s not about winning the argument. It’s not about winning more custody or anything like that. It’s just: how do we keep this 100% child-focused and move forward? If you prompt your agent in a way that those are the results you expect, it usually comes out pretty good.

Don: There’s a use case for everything, I think.

Stefan: What I like about your approach, Rich, is — I think this illustrates it, and also something you were saying before, like working on the curriculum — you don’t just offload it and tell the agent to go do it, or, “Hey, just handle this communication with so-and-so.” You have gone through this process so that you make sure the agent really understands your perspective. But I imagine that process also helps you flesh out that perspective. When you were telling me the other day about having that series of questions for every item, you put a lot of time into that.

Some people might say, how do you really save time if you’re spending all that time putting it in in the first place? But what I really appreciate, and would love to hear your thoughts on, is — how does that maintain your control over the thought and the idea, so that it’s not offloading it; it’s just extending your own ideas and capabilities?

Rich: Yeah. It’s more about refining things than offloading. To answer your question within a question, it saves time because you do it once, and then you know what to expect when you have to do it again. You don’t have to tell it twice — unlike an employee, where if it’s worth saying, it’s worth repeating. You don’t typically have to repeat yourself with AI.

Stefan: But your process — you really retain, you really maintain control over those ideas. Take the example of the curriculum: you go through this process where, for each point, you’ve trained it to ask you so many questions, so that you really flesh out an understanding of it, so that it’s really your idea. Talk about why you design it that way. What’s the benefit of doing it that way? Do you really feel like it’s still your idea and you control it?

Rich: Sure. There’s a quote I keep coming back to recently — I don’t want to paraphrase it or misquote it, but it’s: “Unspoken concerns are premeditated disappointments.” It’s a way of saying: if you don’t communicate every concern you have, every limitation you have, you’re just setting yourself and the other involved parties up for failure.

Rich: I would challenge the professors to say: use AI — absolutely use it, because it is a force multiplier. But be prepared, when you submit that final assessment, whether it’s a project or an exam or whatever, to tell us how you developed it. What questions you had. What obstacles you faced. What you had to do differently — you thought it was going to go this way, and then you had to change direction. Instead of focusing on preventing the use of AI at all, maybe you need some kind of counter, to say, “I had 14 different iterations. I had 87 iterations before I got to this final thing I turned in.” Because that’s where the work is — how many times did you have to iterate to get to the final project. Does that answer your question?

Anand: Yeah, that definitely helps. What you’re getting at is something we’ve talked a little bit about: we’ve got to find a better way to assess student work. It can’t just be “look at the final product.” What you’re really getting at is it’s got to be an interrogation of that process — how did they get there?

Rich: It kind of flips the coin, doesn’t it? Because if the teacher is giving the same student every assignment or every exam and then checking it against an answer key, and that’s how they do their work every day — well, that’s the laziness you’re trying to prevent in the student, isn’t it? What if, instead, you’re having conversations with the students and saying, “How did you come to this conclusion?” or, “In getting to this conclusion, what presumptions did you have to challenge?”

Subscribe now

Chapter 4: Why Better AI Use Starts with Better Questions

Anand: That’s all active learning. Going from a system that largely trains everyone the same way with the same stuff and then gives everyone the same assessment, or similar assessment, to see who conquered it — it has limited use, and it’s going to be hard to move beyond that. But I’m wondering, with your summer program — I’m sure you’re not going to be giving papers and tests, or too many tests. What are some of the, to use a fancy word, instructional modalities or approaches you’re going to take to working with the students? How are they going to learn, demonstrate what they learned, and integrate, I assume, some AI into the projects they’re working on? What approach are you taking for the summer?

Rich: Sure. To give you some context, the eight-week leadership course is divided into four phases. Each week is two four-hour sessions, probably in the evening, but we’re going to try to feel out what works best for everyone involved. Maybe it’ll end up being mornings — I don’t know.

Rich: The first phase is self-awareness and self-leadership. If you can’t lead yourself, how can you lead others? So before they ever set foot in the classroom, we’re going to give them a number of assessments — MBTI, the Big Five. I’m a big fan of the CliftonStrengths test. We’re going to give them all those assessments before they ever come to us. Then they’re going to be divided into five teams, and those teams are going to be structured around conflict. We want them to be in teams where they’re going to have conflict.

We’re going to put them in groups based on people we think they may not get along with, based on their different personalities and how they look at the world. We’re not even going to talk about those assessments we gave them until the second week. First, we’re going to challenge them to produce things as a team and let that conflict come to the surface, and then we’ll talk to them about why they’re dealing with these conflicts in the second week.

The second phase is all about good and bad leadership — what it looks like. Case studies throughout history. We’re going to meet on a battlefield in the Fredericksburg area and do a battlefield study: what decisions had to be made, the decision-making process, how it looks in real life. The third phase is developing and leading teams. The fourth phase — the one I’m most excited about — is our community leadership project. Back in phase two, they’re going to start talking about what they want their leadership project to look like, because as a team, all five teams are going to go find a problem in their community and try to have a positive impact on it. You’re not going to cure cancer. You’re not going to end hunger. But you can find a problem in your local community and try to leave it better than you found it.

Stefan: I mean, that sounds incredible. They should put that in schools. Seriously. That kind of project-based learning is exactly what students need, because not only does it better motivate them — they see they actually have an impact, they’re doing something, it’s not just arbitrary, not just made up from exercises that don’t have a real-world impact — I think you’re going to find the students are going to rise to the challenge. I can’t wait to hear how that works. I definitely want to be in touch with you afterwards to find out how it goes. Maybe we can have you back to talk about that experience with it.

Rich: Love to.

Anand: One thing we were talking about a little when I met with you — you weren’t really designing this program to be AI-intensive, but it seems like it really prepares those students for the use of AI, or for an AI world. Can you speak to that a little? How do you think it prepares them for a world where they’re going to have to use AI, and AI is changing so many of their jobs?

Rich: Yeah. So maybe what you’re speaking to is — I think you and I discussed — the smartest minds in our country right now are talking about a future where in three to five years, two-thirds of all white-collar jobs may be eliminated. And I think a lot of blue-collar jobs are going to go with it. I think they just think of white-collar jobs as being easier to automate, and they’re in the white-collar sector, so they concern themselves with white-collar business. But I think blue-collar jobs are right behind them.

What does that world look like? I think we’re going to move away from the current environment, where you graduate college, you try to find a career you can settle into for 45 years and make a long haul of it. I think we’re moving toward an economy of problem-solving. If it can be automated, it will be automated. So if you want to do something with your life, if you want to find value, you’re going to have to go find a problem to solve. Our SLA is structured around finding problems to solve and going out and leading people to fix those problems.

Chapter 5: Leadership

Don: You mention a lot of going to the battlefield and thinking about some great historical military examples to highlight this. I think some people in society — especially those who haven’t had any military background or family members in the military — kind of view that as separate from leadership in the rest of society. Obviously it’s not. But I’d love to hear you speak a little about that. How is it that the leadership skills, and learning about it from those that were in the military, veterans, or thinking about military experiences — how can we draw on those and have students draw on that, even if they’re not going to go into the military? What are the similarities between the two?

Rich: I think military leadership kind of gets a bad rap, in that it’s viewed as if there’s this commander or drill instructor who stands on a soapbox and barks down at people, gets in their face and yells at them and tells them how to do things — when in reality, the leadership I experienced in the military was all about serving those at the bottom. If I’m a corporal and I’m in charge of lance corporals and below, how do I best enable them to be better at their job and to achieve the mission?

One of the things I think industry has yet to learn from the military, especially the infantry, is — we’re really good at training everyone below you to be able to immediately step into your position and take over your job, because you may be the first one to get shot when things hit the fan. Sorry, I need to consider your audience. I’ll try to censor myself. We’re really good at that. But I think in industry, it’s much easier to find what I’d consider a gatekeeper — “I need to be the only one who can do my job to the best of their ability, and I need to be indispensable, or I’ll be easily replaced.“ I would challenge industry to say: look for the people who are actively trying to replace themselves with their subordinates. Those are the people you need to promote, because those are the people making sure everyone is capable of progressing. Military leadership at its heart is: leaders eat last. I need to take care of everyone in my charge and make sure they’re best enabled to achieve whatever the mission is today.

Stefan: One thing I’ve been thinking about a lot lately is that education can be very self-referential. I can create a content goal or a skills goal, and then we can see how many students met it — the whole system can kind of reference itself. In industry, you can’t really do that, because you have to compete against other companies. In the military, the stakes are even higher — they’re more existential in many ways. Now that you’ve been in the military and you have two of your own businesses, what are some things you think education can learn from either or both — business or the military — in terms of how we should prepare students for the future? I’m sure you reflect on what your own daughter is learning at school and how she’s learning, and you might reflect on that a bit.

Rich: That’s a big question. I think education is always going to be kind of handicapped, because the timeline it takes to get from what industry needs, and what industry is doing well, into the curriculum — and then to prepare students for that — especially as things change quickly… I think we talked about the OODA loop. It was Colonel Boyd who created the OODA loop, an acronym for Observe, Orient, Decide, and Act. He was an Air Force colonel teaching dogfights. If you have two airplanes in the air actively trying to defeat one another, the one who can observe the enemy, orient, decide to pull the trigger, and pull the trigger to eliminate their foe as quickly as possible — the one who can do it first wins.

He then applied that and said: we do thousands of these OODA loops every day in every part of our lives. The better you get at executing that loop, and at preventing people from interrupting yours, or the better you get at interrupting your opponent’s, the more effective you’re going to be. The problem we’re all facing with AI is that loop is getting tighter and tighter, and it’s difficult to keep up with.

Subscribe now

Chapter 6: Learning with AI

Rich: Industry is going to change very rapidly over the next couple of years, and I don’t know how education can keep up with it. I don’t know if it’s possible to keep up with it in its current form. I’d pose the question back to you: how do you keep up with something so advanced? Government can’t do it in any way, shape, or form, and school is regulated by government. How do you take the lead and stay in front of that ship? I don’t know.

Stefan: It even goes back to your summer program — people, whether through the nonprofit or for-profit sector, offering students ways to learn outside of school. A lot of students are taking advantage of those opportunities in other ways. I think students have started just learning things online. I’ve had a couple of students where I am now who learned a lot about AI — they taught themselves both the math behind some of the models and some computer-science things. That’s probably not the average student, but there are a lot of opportunities to learn more than you need.

I don’t think schools are going to be able to provide all the instruction. I don’t think most faculty are going to learn that quickly at either the K–12 or university level. But to really make those resources available to students — encourage them to take classes on Coursera, or even just find things online where they can learn… I think the one thing schools can do that they’re not doing — that they’re almost afraid to do — is to tell students, “The world’s changing a lot. We can’t teach you everything you need to know. You’re going to have to take more responsibility.” Schools, especially K–12, are so protective of the kids, trying to make them do things, which is okay, but that’s not really going to be enough. I think schools could really empower students more than they are. Sometimes they’re almost disempowering them, saying, “Just do exactly what we tell you and you’re going to be set.” I don’t know how anyone can say that.

Rich: Yeah. I think if the schools have any duty to the student, it’s to teach root principles and instill a sense of judgment. You can create anything with AI, and you can even ask it for five solutions to the same problem — but knowing which of those solutions is the best, that’s a skill that’s going to have to be taught, or it’s useless.

Stefan: I think that’s an incredibly important point, Rich. That’s something we’ve talked a bit about in other interviews and episodes, but most people in education don’t get that. It’s changing — there are some who are still resistant and thinking we can just hold off AI, and that’s not realistic, obviously.

Don: Yeah. I mean — why would they want to?

Stefan: But for many others, to appreciate this role of judgment… I think we’ve always thought we were teaching some of that, but it’s back to your example of the OODA loop — it’s happening so much faster, and that level of judgment has got to be so much more specific and more sophisticated than ever before. When I think about evaluating somebody’s arguments — if somebody made a claim, if we’re having a discussion about politics, the arguments we would have made a generation ago would probably be fairly basic. Now we have access to so much more information; things are far more sophisticated. Not only do I need to understand that argument, I need to be able to judge it in a more sophisticated way, evaluate it in a more sophisticated way. I think that’s where we’re falling behind — we’re still looking at it through the old lens.

Rich: Are you suggesting that critical thinking is important to know? It just might be.

Anand: This is something we’ve been on a soapbox about for a long time, because we both have a debate background. If you don’t have critical thinking, I don’t know how you survive in this society. You’re just going to be taken advantage of left and right.

Rich: I really appreciate Elon Musk’s stance on this whole subject — that he started very pessimistic, very doomer, saying this is not something we are ready for, not something we should be actively developing, and we need to pump the brakes. And then it wasn’t long after that that he came out and said, “You know, the genie’s out of the bottle. You can either participate or spectate. So I’m choosing to participate.” And in that — I’m obviously inserting my own thought process here — he’s saying the genie is out of the bottle. It’s up to us to decide whether that genie looks like Robin Williams or more like Jafar. The duty of our educators right now is: how do we make sure we’re producing students who want to create things that look like Robin Williams and not Jafar?

Subscribe now

Chapter 7: Process Over Product: Rethinking Student Assessment

Anand: When you’re thinking about your own — I love a good Aladdin reference; you can tell somebody has a kid, they’ve watched that movie a number of times — when you’re thinking about your own daughter…

Rich: The OG Aladdin, right? Not the live.

Anand: Yeah. What do you want your daughter to develop as far as these skill sets? That’s the hard part. I have kids in public schools too, and they have to play the game — there are still things you have to do — but I’m also like, “Okay, this is what you really need. This is what you really need to be able to do.” How are you guiding that for your own child?

Rich: Yeah. Again, spectate or participate — and I’m choosing to participate as much as possible. We both have full schedules at the moment, and so we get a little bit of time in the evening and in the mornings. I’m very grateful that I’m fortunate enough to be a parent who can take my daughter to school and pick her up from school and have conversations around that. This summer I’m planning to give her her own AI agent, and sit down with her and — as we talked about over coffee — set it up in a way that not only doesn’t spend a ton of money on tokens, but doing all the work to dictate or direct the personality and the values of that agent to produce useful and meaningful things. We’re going to sit down this summer shoulder-to-shoulder and build her an AI agent — or ecosystem, or whatever it looks like. We’re going to learn a lot of it together, I’m sure, because by then things will be different and I’ll have to relearn things. Active participation is the best thing you can do as a parent, and that’s what I’m choosing to do.

Stefan: I’d love to be a fly on the wall on her first day back to school, and you know, sometimes a teacher says, “Hey, everybody, say what you did over the summer.” And your daughter says, “My dad built me an AI agent to do all my schoolwork.”

Rich: Well, that’s okay. Remember — I’m a military veteran and a bodyguard by trade. My biggest concern going into kindergarten was that she was going to tell her teacher that I hide guns and money in the walls. So I think we’ve progressed a little further. I think this is a safer topic for her to talk to her teacher about. That’s fine. Yeah, I think so. Maybe.

Anand: What I love about that example, though, is it’s not just “Yeah, kids are going to have AI, let them have AI.” It’s playing an active role in then guiding what your daughter’s going to do with AI, helping her build it, talking to her about it, and letting her see your process for building it out. One thing you told me about, with what you’ve done with your OpenClaw, is training it on certain materials so that it gains a particular perspective. Tell us a little bit about that. What went into that? Why do you think it’s important for that agent to have a particular perspective? This is something we’ve talked a lot about — what we call AI pluralism — this idea that every human obviously has a different perspective, AI is going to have a different perspective. This really resonated with me. Tell us a little bit about what went into it and why you did it.

Rich: Yeah, I’ve kind of created my own little internal LLM for my agents. I’ve downloaded ebooks and PDFs of what I think is important work, and the things that shape my mindset — therefore I wanted to shape the mindset of my agents who work for me. That’s a whole lot of books on leadership, from Jocko Willink, Jordan Peterson, military generals throughout history. The very foundation was the Stoics — my agent is called Marcus, because Marcus Aurelius was a layup for me — and the disciples.

You can see it in a lot of the content we produce — it’ll very often reference the Stoics, or the disciples, or Jocko Willink. Because that’s the information that has developed my brain, it’s also the information I wanted to put into my agent. I think that revolves around a conversation we had — if the next iPhone doesn’t have an agentic AI platform built into it that is shaped and developed around the information they’ve already gathered on you as their user, they’ve missed the off-ramp completely. These big data users have gathered so much information — as we talked about, if you have Tinder installed on your phone for a week, how much information they’ve gathered on you, about when you’re lonely and when to serve you different types of ads — if they don’t start using that to actively develop your own agent for you, that’s a huge missed opportunity.

Anand: So are you thinking about helping your daughter to train her agent in a similar way — to take some of her favorite books or materials? What I love about that is — it would be highly egocentric of me to think that my daughter wants to learn from the Stoics, but…

Rich: Yeah, I think there’s some of it in there. I want her to have the same idea as I do.

Stefan: I mean, hers might be train on Diary of a Wimpy Kid or something — whatever she feels is important. I love that idea of imprinting yourself on it so that it better represents you as the user.

Rich: Absolutely. It’s kind of like the quote, “If you don’t define success for yourself, others will define it for you.” In the same way, if you don’t tell your agent how you want it to think and what you value, you’re waiting for these companies to do it for you — and I think that’s the wrong approach.

Anand: Yeah — if you don’t imprint yourself, you’re going to be imprinted, right? It’ll gather, like you say, potentially all that information they have on you, and most people don’t understand the totality of it — what they’ve gathered on you for decades. That still may shape, or kind of misunderstand you, at least according to you. Maybe they understand you better than you understand yourself, so to speak. But you want to have an opportunity to influence that — that’s one thing.

Subscribe now

Chapter 8: The Service Leadership Academy Model

Stefan: The two other things that stand out to me, as I’ve listened to you talk: one is that now you need to be a continuous learner. It seems like you’ve taught yourself all this stuff on your own. Agents are a new thing; OpenClaw is a new thing. You’ve learned a lot of this on your own. Can you talk about how you went about learning how to build your own agent and use OpenClaw?

Rich: Well, to give a little more context — I think I told this already — I was a freshman in high school reading Rich Dad Poor Dad and books about starting your own business. I was 100% drinking the Robert Kiyosaki Kool-Aid, and I wanted to own the cash-printing machine. I didn’t want to be a part of someone else’s machine. Then as a sophomore, a couple of guys flew airplanes into the World Trade Center, and I could not join someone else’s machine fast enough. I enlisted in the Marine Corps the day after my 17th birthday, and I left for boot camp four days after high school graduation.

I spent eight years in the Marine Corps. I was diagnosed with traumatic brain injury as a result of an IED about three and a half years before I got out of the Marine Corps. I started my first business two years before I got out. They had told me at the time that I was doing well teaching student lieutenants, and so they were going to allow me to continue to do that, but they didn’t see me going back to the fleet and leading troops in combat with my brain injury. So I had to make a huge adjustment. I had to pivot, and I started having ideas about being a business owner again.

Backing up a little — as a freshman, I think I probably had a 4.0 GPA. Then I joined the Marine Corps, and I knew all I had to do was get a certificate, get that diploma, and I’d be fine. So I spent all my time from that point forward focusing on sports and girls. About the only thing I have in common with Patrick Bet-David, other than some of our values, is that we both graduated high school with a 1.8 GPA. I have gone back to school — I’ve done the community college route, and I used a bunch of my GI Bill to go to professional schools like Executive Security International out in Colorado. So I definitely view the importance of knowledge and education. I don’t necessarily agree with how it’s being delivered nowadays.

Rich: By the time I adopted OpenClaw — to get back to the root of your question — there was already way too much information on YouTube. I get most of the information I need to do the things I want to do either from YouTube or from the AI itself. You can learn a lot about how to use AI by asking AI how to use AI, and having it teach you. Now, it’s in a vacuum — it’s going to tell you based on what else is out there. But I think the combination of the two for me, whether it’s YouTube or X or asking AI, has done well to teach me what I need to know. And I tell you what — refining AI production by having a different AI check the work of the predominant AI you’re using produces pretty good results.

Don: It does. Yeah. It’s just really incredible how AI can teach you so many things. There are so many ways to learn. It’s kind of a theme I’ve picked up on. Sometimes it doesn’t go over well with educators, but I’ll say: ask your students what some ways they learn are besides in the classroom. Even just regular school math or whatever — they can get it from the teacher, they can go on YouTube, they can ask the AI; some kids have tutors. We get upset when kids aren’t paying attention in class, but it’s not the only place they can learn content. In some instances, it’s not the easiest place for them to learn. There are so many opportunities to learn and grow and build businesses. It can be a little scary — you’ve got people losing jobs, and you need to be aware of hallucinations — but you can do the same thing with an AI agent that you can with a student, and say, “Show me your work. Show me the citation.”

Anand: This sense of being able to learn in other places — we’ve got to lean into that. So much of the traditional educational approach is that the classroom is this sacred place. Stefan and I are educators, so we have drunk that Kool-Aid before — but we also, I think, realized as debate coaches, and from our experience in debate, that I learned a lot more from debate than I ever did in a classroom. Because I was motivated. I wanted to win that debate. I wanted to make sure that I didn’t look like an idiot when I was in the middle of a debate. So you find other ways to learn. We’ve got to find a way for students to be motivated to learn like that.

Rich, if you could speak to most — let’s say K–12 educators or higher ed — what would you tell educators about how do they make that pivot? It’s not easy when you’ve been doing something for a long time the same way and you realize, “Hey, things have to change.” I’m not sure how to change. What would you say they should start with?

Rich: Well — you brought up debate. One of my favorite debaters is Dennis Prager. I don’t know if you’re familiar. He’s a conservative Jewish debater. He owns an organization called Prager University; the YouTube is PragerU. He will tell you that the reason he often gets invited to debate in circles that a lot of other conservatives don’t — or anyone who challenges the view of the people in that circle — is because he doesn’t seek to win the debate, necessarily. Anytime he’s on a debate stage, he certainly doesn’t look for gotcha moments, but he doesn’t necessarily look to win the argument so much as to better understand the person he’s debating against.

If he can better understand what’s brought them to the conclusions they have, and what’s driving their thought processes — whether he wins or loses, that’s a better state at the end of the debate. With every single student, if you seek to better understand what’s driving them to where they are, how they got to the conclusions they have, and then use that to better influence them and better educate them, I think you’re going to be light-years ahead of where you started.

Stefan: And talk about creating a space where the students are going to feel valued. That’s a great approach.

I’ve really enjoyed this conversation, just thinking so far outside the box. You talk about how your educational pathway was not traditional — some educators may look down on it and say, “Well, this person graduated from high school with less than a 2.0” — but you’re talking about project-based learning, you’re having educational programs for students, you talk about your own learning path. It’s a lifelong learning, and having agency, and developing businesses. A lot of students who go through the traditional pathways are not going to end up there, and they’re not going to end up with the skills they need to succeed — not just in this new world, but to some degree as in previous worlds, too.

Rich: Maybe — maybe as we get students through the program, and we’ve graduated a handful of them, I can start branding myself as the heretical educator.

Stefan: That’s awesome. That’s awesome.

Subscribe now

Chapter 9: Preparing Students for an AI-Driven Economy of Problem Solving

Don: I know — we met on LinkedIn, and that’s where a lot of this conversation happens. It is project-based learning, and he’s doing this pretty well, but he’s just working so hard to get schools to do it on a really limited basis — like a special thing in the class — and you’ve come up with it, and you’ve taught yourself all this stuff, and you designed the program and thought through all the iterations of it, like you said, with an AI. It’s really awesome. It’s just a demonstration of not just doing well, but — like you said, you could take this powerful AI and you could basically do good or bad with it. You’re doing good with it.

Rich: I appreciate that.

Stefan: I’m really excited about the summer program, and I definitely want to follow up with you to find out how it’s going, because that’s exactly the kind of education students need. When we were talking a little earlier in the conversation about the problem with more traditional approaches — somebody going and getting a degree takes years, and even a course will take so long when you need the skills now. We have to find ways for education to adapt, or maybe for traditional education to get out of the way for people to be able to develop those skills. If you waited and said, “Hey, I heard about OpenClaw — I think I’ll sign up for a class next fall to learn about OpenClaw,” and then maybe by December there would have been a class about it, at that point OpenClaw is not even going to be around. Things will change so rapidly. It’s just laughable to think about a traditional educational process working with something that’s changing that quickly.

Anand: There are some areas — I still want to have a doctor that’s gone through a lot of training. You’re right, root principles matter. But for most of us, when we’re thinking about these skill sets, we’ve got to find ways for students to be motivated to go and learn it right away. It’s got to be just in time, and it’s going to come down to their motivation. That’s one thing I love about what you’ve outlined — giving students that agency and that motivation is key. I think you have a great way for them to develop that.

Rich: Well, I want you to think about the words you just chose. You said, “I want a doctor that’s had a lot of training.” And we’re getting very close to having AI models that have far more training than any human can get in a 20-year education. So — I think we agreed — the future is equal parts exciting and terrifying. It’s going to be a hell of a ride.

Stefan: That it is. That it is.

Anand: If you don’t mind, one last question I want to ask you about: I hear often from people that they’re — especially living in Fredericksburg — resistant to AI because of data centers. They don’t want AI to displace their jobs. They just want to put up roadblocks. What do you say to them? How do you get people to appreciate that maybe we need to take a different approach?

Rich: I would say: keep up. The data center problem — by the time we’re done arguing about it, Elon will have data centers in space, and it won’t be something to argue about anymore. It was the same thing we saw in the ‘90s and early 2000s, when Congress couldn’t come to an agreement about whether we should be cloning sheep, or using stem cells from fetal material to study DNA. They couldn’t finish having the argument and come to a consensus before science came back and said, “Yeah, we don’t need that anymore.” So the thing is — you can either keep up, or get left in the dust, because there’s really no alternative.

Stefan: It’s hard to find something more powerful to end on. What a great — you should open a keynote at a university. That can be your opening line. Very serious.

Rich: I’ll do that one on pay-per-view, you know.

Anand: Rich, what Stefan’s saying is — you’re among kindred spirits here. We’re all on the same page on this, and our worry is that people aren’t going to keep up, and we’re mostly worried about our students. We’ve got to find a way for them to keep up and be given these opportunities. That’s one of the reasons I’m so excited about your program this summer — because even if you never even mention AI, those students are going to be better prepared for an AI world than most other student experiences they would have. So I applaud that, and I encourage you to keep building that.

Stefan: Reverse education. They’re being told, like, “Don’t keep up, don’t do this” — it’s kind of the opposite of what they mean.

Chapter 10: What Education Can Learn from Military Leadership

Rich: I don’t know how much more time you have, but I want to give you a real quick, bigger view of what we are doing. The SLA — the Service Leadership Academy — is one aspect of what the Guideon Foundation is doing. The Ronin Leadership Activation Program — there’s kind of a huge big picture here.

Rich: Veterans are committing suicide at a rate of 22 per day. There’s another whole number of veterans who are going homeless every day. It’s a misconception that that’s all due to combat trauma and PTSD — that’s a misconception, because most of them haven’t been seeing combat. So why are they removing themselves from society? I think it’s because we’re not preparing them for the huge transition that takes place when you go from wearing the uniform and having a certain amount of reverence — but also having a built-in purpose and identity that goes along with that uniform. You never question your purpose or your tribe while you’re in uniform. Then to separate from that tribe and have to struggle to figure out who you are and what your worth is, because you’ve been tying your entire self-worth up in what you do for a living for the last four to 30 years — that’s a very difficult transition to make.

Rich: I only recently realized that veterans and stay-at-home moms who are empty-nesters have a very strong connection. They’ve tied their entire identity to this one job they do for the last 20 years, and now that job doesn’t exist. So how do you find value in yourself?

I say all this because — it’s kind of like the LA wildfires. The dumbest thing about the LA wildfires was that we had these wildfires that were wrecking people’s livelihoods, and we had this huge reservoir of water to put out those fires with — that we refused to touch. We had the problem right there, and we refused to use the solution to solve the problem. More recently, we had 500 tons of food in Africa that was paid for by American taxpayers to feed the poor in Africa, and we destroyed all that food because of bureaucracy. We had the solution, and we chose not to use it to solve the problem.

Rich: Again — we’ve got veterans throughout this country who have millions of dollars and countless hours in training to be leaders. And we have a nation that’s starving for leadership. So why not use those veterans to solve that problem? And if we have two-thirds of all people having to go through the same transition in the next three to five years — where now they’re struggling to find identity because the career they’ve had for 20 years is completely gone because of AI and robotics — well, they’re going to be going through that struggle too. So what the Guideon Foundation is hoping to do is to get veterans in a position to better help the rest of our population, who very soon is going to be saying, “What value do I have in myself if my career doesn’t exist anymore?”

Stefan: That is really powerful, and I think you’re right. It’s a tragedy that veterans are experiencing this, so it’s important to address that. But you’ve tapped into something much bigger. This is going to be a transition that will affect so many people, that if you’re able to work on this, it could benefit so many sectors of society.

Rich: Yeah. Equal parts terrifying and exciting.

Anand: More of an emphasis on terrifying for some of us. But it’s exciting to see you working on this, and it’s really great that you’re giving your time to work in these areas.

Rich: Well, it’s the one way I’ve found to try to be useful.

Stefan: Thank you, Rich. Thanks so much for being part of this. If somebody wants to find out more about the Guideon Foundation, how can they find you or the foundation’s materials?

Rich: Sure. guideonfoundation.com — guideonfoundation, no spaces, no periods, nothing. On Instagram, Facebook, YouTube — all the places.

Anand: Rich, thanks for joining us. We’re going to follow up. We want to hear more about what happens this summer, and as I said before, I want to bring you on campus for some discussions. I think a panel discussion with local business people about how they’re using AI would be so valuable for our students, and to be able to share what you’re doing with the Guideon Foundation. I think that’d be great.

Rich: However and wherever we follow up, I’m already looking forward to it.

Anand: Excellent. Hey, thanks so much.

Rich: Thank you.

Subscribe now

In this post, I want to offer my thoughts on today’s publication of the Pope’s Encyclical on AI.

Today, Pope Leo XIV presented Magnifica Humanitas, his encyclical on safeguarding the human person in the age of artificial intelligence. The document was signed on May 15 and was promulgated today, May 25, 2026. He signed it 135 years to the day after the encyclical that defined the Church's answer to the Industrial Revolution.

It was launched in collaboration with Anthropic co-founder Chris Olah.

I have spent the day reading it closely, especially from the perspective of education (my underlined version is here). This is not a short statement, a technology policy, or a set of classroom guidelines. It is a major institutional document, running to roughly 250 paragraphs across five chapters, that treats AI as a civilizational turning point — one that will reshape work, truth, power, human dignity, democracy, and the formation of young people.

I should say at the outset that I am not approaching this as a religious person, and I am not writing as a Catholic. I draw more from the document’s moral, educational, and social arguments than from its theological claims. Others will read it differently, and that is fine.

But even for secular readers, the encyclical matters because it does something most educational institutions have not yet done: it treats AI with the seriousness of a world-historical event. The Pope is not asking whether students should be allowed to use ChatGPT on essays and what AI traffic lights to put up. He is asking what happens to human beings and education itself when intelligence becomes increasingly artificial, automated, conversational, and ubiquitous.

That is the level of conversation schools should be having. Most are not.

This post is not a summary of the whole document. It is a reading of the encyclical through the lens of education — what it asks of schools, what it gets right, where I think it overreaches, and why its seriousness should further push many educational institutions into action. It is organized around six things I want to say.

First, that the form of the encyclical matters, and the contrast between how the Catholic Church is responding to AI and how most educational institutions are responding to AI is the first thing worth noticing.

Second, that one paragraph of the encyclical — paragraph 99, on what AI is and is not — overreaches in a way the rest of the document does not, and a serious reading has to engage that overreach honestly.

Third, the responsibility the encyclical places on educators, and especially on school leaders, is far heavier than most educators have allowed themselves to feel.

Fourth, the catalog of risks and harms the encyclical names, including those that impact the young, is comprehensive in a way that should make sympathetic readers feel that their worries are not being dismissed.

Fifth, that the criteria the encyclical proposes for evaluating AI — solidarity, the common good, subsidiarity, access, dignity, social justice, integral development — give us a coherent framework for asking whether the systems being built and deployed serve human flourishing or undermine it.

And sixth, what to make of all this — where the document leaves us, and what it leaves us to do.

I will spend the bulk of this post on the first three, because they are the most demanding for educators. The fourth and fifth are shorter surveys. The sixth is a brief conclusion.

Subscribe now

1. The significance of the form, and the contrast with schools

Before I get to what Magnifica Humanitas actually says, I want to spend a moment on what kind of document it is. The form matters here — maybe as much as the content.

Most discussions of AI in schools begin with a question about tools. Should students be allowed to use ChatGPT? Should teachers grade with AI? Should we adopt this platform or that one? These are not trivial questions. But they are not the questions Leo XIV is asking in Magnifica Humanitas, and the gap between his question and ours is the first thing worth noticing.

In paragraph 92, the Pope writes: “This makes it clear that technology is not simply a tool. When it becomes the standard by which everything is judged, it begins to dictate what matters and what can be discarded, reducing creation to an object of exploitation and human beings to mere cogs in a system driven toward ever greater efficiency.”

That is a philosophical claim about what AI is and what it does to the institutions that adopt it without thinking carefully. Technology is not simply a tool, because once it becomes the measure of everything, it starts deciding what counts as success. Efficiency. Throughput. Speed. The things that are easy to measure rise. The things that are hard to measure — judgment, character, friendship, the slow formation of a young person — fall off the dashboard. The Pope is warning that an institution which adopts the technology without first deciding what it exists for will, sooner or later, find itself reorganized around the technology’s priorities rather than its own.

The form of an encyclical

This is what makes the form of the encyclical itself so significant. An encyclical is the highest form of papal teaching short of a formally defined dogma. The word comes from the Greek enkyklios, meaning “circular” or “going around” — these letters were originally circulated to all the bishops of the Church, and through them to the faithful. Today encyclicals are addressed not just to Catholics but, in most modern cases, to “all people of good will.” They are meant to be read, taught, debated, and acted on for decades.

Popes don’t write many of them. Leo XIII issued Rerum Novarum in 1891 on labor and capital. Pius XI issued Quadragesimo Anno in 1931 on the structure of the economic order during the Depression. John XXIII wrote Pacem in Terris in 1963, during the Cold War, at a moment when nuclear annihilation felt imminent. John Paul II wrote Centesimus Annus in 1991, as the Soviet system collapsed. Francis wrote Laudato Si’ in 2015 on climate and our common home. These documents are how the Church marks the moments when it believes a civilizational hinge has turned — moments when humanity is facing what Magnifica Humanitas calls the res novae, the “new things” of an era.

That is the company Magnifica Humanitas is keeping. Leo XIV is explicitly positioning artificial intelligence alongside the industrial revolution, the rise of totalitarianism, the nuclear age, the collapse of the Soviet bloc, and the climate crisis. He is saying, in the most weighty institutional form available to him: this is one of those moments.

It is worth pausing on what the Church is not saying by issuing this document. The Vatican is not betting that AI is a speculative bubble about to burst — the way some economists and skeptics have suggested, comparing the current investment frenzy to the dot-com era. An institution that thinks in centuries does not write encyclicals about fads. Nor is the Church suggesting that AI’s challenges can be resolved with the kind of red-yellow-green policy traffic light that has become depressingly common in school AI guidance documents — green-light tasks where AI is fine, yellow-light tasks where it requires permission, red-light tasks where it is forbidden. A 50,000-word theological treatise grounded in 130 years of social doctrine is not what you produce when you believe a chart will suffice. The form of the document is itself a claim: this challenge is structural, civilizational, anthropological, and it requires the kind of sustained moral reasoning that grids and rubrics cannot provide.

Whatever you think of the Catholic Church — and as I said at the top, I’m not approaching this from a religious place — this is an institution that thinks on the order of centuries. It does not panic. It does not chase news cycles. When it convenes itself to issue an encyclical, it has decided that what is happening is structural, not episodic. Magnifica Humanitas tells us that the Vatican has read the room and concluded that AI is reshaping the human condition itself — the nature of work, of truth, of communication, of power, of war, of what it means to be a person — and that the Church must address it with the same seriousness it once brought to industrial capitalism.

Subscribe now

Now contrast that with how schools are responding

This is where I want to sit for a moment, because the contrast is jarring.

The Catholic Church — an institution with roughly 1.4 billion members, several layers of bureaucracy, and a deliberation cycle measured in generations — has produced a book-length, theologically grounded, philosophically rigorous, internationally translated statement on AI within roughly three years of ChatGPT’s release. The Pope is treating this as a defining issue of his pontificate.

Meanwhile, most schools, districts, and universities have produced: an acceptable use policy. Maybe a three-pager. Maybe a committee that met half a dozen times. Maybe a workshop in August that half the faculty skipped. Many have produced nothing at all and are quietly hoping the problem will resolve itself.

What would it look like if educational institutions responded with the seriousness an encyclical represents?

It would look like a head of school, a superintendent, or a university president saying: this is the defining issue of my tenure, and I am going to organize the institution around it. It would look like a foundational document — not a policy, a document — produced by the institution’s most serious thinkers, articulating a vision of what education is for in a world where machines can do much of the cognitive work. It would look like that document being translated into the institution’s various “languages” (board strategy, curriculum, hiring, faculty development, parent communications, student-facing materials) and revisited every year. It would look like the equivalent of bishops being convened — department chairs, division heads, lead teachers — and being held accountable for implementation. It would look like a willingness to say that some of what we have done for the last fifty years no longer serves students and must be rebuilt.

That is the scale of the response Magnifica Humanitas represents. A 1.4-billion-member global institution that thinks in centuries took three years to produce a substantive, organizing vision. Most schools, with maybe 500 students and one building, have not produced one in three years either — but in their case it is not because the work is hard. It is because no one has decided it matters.

Subscribe now

And the decision is not morally neutral

There is one more move the encyclical makes early on that I want to flag, because it forecloses the most common educator escape hatch.

The standard line, in faculty meetings and parent forums and ed-tech vendor pitches alike, is some version of the tool is neutral, it depends how you use it. AI is just a tool. A calculator was once a tool. The internet was once a tool. We learned to use those well, we will learn to use this one well, and the question is just about good practice — appropriate use, responsible use, ethical use. Pick your adjective.

Paragraph 104 refuses this framing in language educators in particular should sit with.

“We cannot consider AI to be morally neutral. In reality, every technical tool embodies choices and priorities through what it measures, ignores and optimizes, and how it classifies people and situations. If a system is designed or used in a way that treats some lives as less worthy, or excludes them without the possibility of appeal, then it is not merely a tool ‘to be used well,’ since it has already introduced criteria that contradict the inalienable dignity of the human person. For this reason, ethical discernment cannot be limited to asking whether we are using a system for good or bad purposes; it must also examine how that system is designed and what vision of the human person and society is embedded in the data and models that guide it.”

Read that twice. The Pope is saying that the act of choosing which AI to adopt and how to adopt it is itself a moral act, before any question of how it is used arises. The system has already decided what to measure and what to ignore. It has already decided how to classify the students it sees. It has already decided whose work it considers valid, whose voice it considers fluent, whose answers it considers right. Those decisions are baked in. By the time the tool is in front of a child, the moral choices have already been made — by the engineers, by the product managers, by the executives, by the investors, by the training data, by the optimization targets. The teacher who chooses to deploy it is not choosing a neutral instrument. The teacher is ratifying every choice that went into it.

This is what the encyclical is asking educators to take seriously. Not just how should we use AI in the classroom but what is embedded in this particular AI, and do we endorse it. What does it measure when it grades an essay, and does that match what we believe writing is for? What does it ignore when it answers a student’s question, and does the ignoring distort the student’s understanding? Who does it classify as a strong student, and who as a weak one, and on what basis? What vision of a literate, thoughtful, ethical young person is the system optimizing toward — and is that the vision we hold?

These are not questions a generic acceptable use policy can answer. They require institutions, and the educators within them, to take ethical responsibility for the systems they deploy and the decisions they make around deployment and instruction with and about AI, in the same way they would take responsibility for any other curricular choice.

Subscribe now

Alignment is not enough if a few define the values

Three paragraphs later, the encyclical makes the move that I think completes the argument and that schools need to sit with most carefully. If paragraph 104 says the tool is not neutral, paragraph 107 says and you cannot leave the values question to the people building the tool.

“We cannot be satisfied with merely calling for the moralization of machines — the so-called ‘alignment’ of AI with human values — without also having the courage to insist on a further condition: the possibility of openly discussing the ethical frameworks involved and subjecting them to shared standards of social justice. Otherwise, those who control AI will impose their own moral vision, which will become the invisible infrastructure of these systems. A more moral AI is not enough if that morality is determined by a few.”

This is the Pope intervening directly in one of the central technical conversations of the AI industry. “Alignment” — the project of getting AI systems to act in accordance with human values — is the dominant framing inside the labs themselves. Anthropic, OpenAI, DeepMind, and others have whole research teams dedicated to it, and the public conversation often treats alignment as the moral safeguard that justifies continued capability scaling. Yes, the systems are powerful, but we are working on alignment.

Leo XIV’s response is structurally devastating. He does not reject alignment. He says alignment is necessary but radically insufficient — because the prior question is whose values the system is being aligned to, and that question cannot be answered by the engineers, executives, and investors who happen to be doing the aligning. A more moral AI is not enough if that morality is determined by a few. The values themselves have to be the product of broad democratic deliberation, subjected to shared standards of social justice, openly discussed by the communities the systems will shape. Otherwise alignment is just a name for the moral vision of a few becoming the invisible infrastructure of everyone’s lives.

Now translate this onto education.

Schools are, right now, in the middle of their own alignment problem. They are deploying AI tools whose values are baked in. Tools that optimize for certain outputs and ignore others. Tools that classify some kinds of student work as strong and others as weak. Tools that embed assumptions about what writing is for, what learning looks like, what counts as a correct answer, what counts as cheating, what kind of feedback is helpful, what kind of student is succeeding. Every one of those decisions is a values decision. And in most schools, those decisions are being made by exactly the actors the encyclical warns about — the ed-tech vendors who built the tool, the engineers who trained the model, the executives who set the product roadmap, the investors who funded the company, and a small number of administrators who decided to adopt the platform.

The teachers were not consulted. The students were not consulted. The parents were not consulted. The broader community was not consulted. The educational equivalent of “shared standards of social justice” — what we believe about how children should be formed, what we believe a literate adult should be able to do, what we believe about the relationship between effort and learning, what we believe about the role of struggle in growth — was not opened up to deliberation. The values were imported, pre-aligned, ready to deploy. By the time the tool reaches the classroom, the moral choices have already been made.

Just as the encyclical says we need to figure out our values before aligning machines to them, schools need to figure out their educational objectives before adopting tools that quietly do the figuring for them. What is the purpose of a writing assignment? Why do we ask students to do math by hand? What does it mean to read well? What valuue(s) do math and reading serve? What kind of citizen are we trying to form? What does it mean for a student to grow? These are not technology questions. They are values questions. And they have to be answered in advance, by the community of people responsible for educating young human beings, through the kind of open, contested, deliberate conversation the encyclical insists on.

You cannot adopt the AI tool first and then figure out what you believe about education. The tool will tell you what to believe — in a thousand quiet decisions about what to measure, what to ignore, what to reward, what to call success. By the time you notice, the values will be the infrastructure. The Pope is saying that this is the thing we cannot let happen at the civilizational level. The educator’s job is to make sure it does not happen at the institutional level either.

The Pope thinks AI is reshaping what it means to be human. Your kid’s school thinks AI is mostly a cheating problem.

That gap is the real story.

Subscribe now

2. Where the encyclical overreaches: the anthropological claims in paragraph 99

Before I go further into what Magnifica Humanitas gets right, I want to spend a section on where I think it overreaches. Reading any document — even a papal one — without noticing its weak points is not actually taking it seriously. It is just deferring to it. And the encyclical itself, repeatedly, invites dialogue with the human sciences. So in that spirit, I want to engage one paragraph as a critic rather than as a sympathizer.

Paragraph 99 is the place where the encyclical moves from moral reasoning, where it is on extremely solid ground, into philosophy of mind, where the ground is much less settled than the prose suggests. The relevant passage:

“These systems merely imitate certain functions of human intelligence... So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience... They may imitate language, behavior and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce.”

Read this carefully. The Pope is not making moral assertions. He is making metaphysical and empirical ones. AI does not feel pain. AI does not have experiences. AI does not understand what it produces. AI does not have a moral conscience. These are claims about what is true of these systems, not about what we should do with them. And they are claims that are, at this moment, among the most contested questions in serious thought.

The position the encyclical adopts — that subjective experience requires biological substrate, embodiment, and relational maturation — is one view, held by serious thinkers including the late John Searle and many phenomenologists in the Heideggerian tradition. But it is contested by equally serious thinkers — David Chalmers, Susan Schneider, Eric Schwitzgebel, and a growing number of cognitive scientists — who argue that consciousness may be substrate-independent, that sufficiently complex information-processing systems may instantiate experience whether or not they have biological bodies, and that we currently have no reliable way to verify which view is correct. Anthropic has taken the question seriously enough to commission active research on the moral status of its models. Whether or not you find that research convincing, the fact that it is being conducted should tell you the question is not settled. The same is true of valence (whether systems can have states good or bad for the system itself), of embodiment (the line between “language processor” and “embodied agent” is being actively dissolved by current robotics research), and of the claim that human growth-through-relationship is non-computational (which is one position in philosophy of mind, not a consensus view).

Alex Wissner-Gross has pointed out that these thoughts won’t necessarily be on the right side of history, as Budhhist orders in South Korea are ordaining embodied AIs as monks. Ben Goertzel has also stated the Pope probably didn’t go far enough, pointing to an emerging global AI consciousness and some previous Catholic scholarship/teaching that went in that direction.

Why this matters

I am not raising these objections to dismiss the encyclical’s broader argument. I think Leo XIV is right that we should be very careful about anthropomorphizing AI, very careful about substituting machine interaction for human relationship, and very careful about treating computational outputs as morally equivalent to human judgment. These are wise practical commitments and I share them.

But there is a difference between practical wisdom under uncertainty and metaphysical certainty masquerading as practical wisdom. The encyclical, in paragraph 99, makes the second move. It tells us, as a matter of fact, what AI is and is not — at a moment when serious people who have spent their careers studying these questions are deeply uncertain about the answers.

This matters for two reasons. First, we are entering a period of radical experimentation on radically new technology, and we should not pretend we know what these systems are. We are training systems with more parameters than we can fully understand, on data we cannot fully audit, using techniques we cannot fully explain, observing capabilities we did not predict. To declare, in this context, that these systems definitely do not have experiences, definitely do not understand, definitely cannot mature, is to make claims we do not have the epistemic standing to make. We do not know. Not knowing is the most honest place a thinker can stand right now.

Second, the moral weight of the encyclical’s good arguments — about dignity, solidarity, the common good, the dangers of profit-idolatry, the duty to remain human — does not depend on the metaphysical claims in paragraph 99. Even if future AI systems do have some form of experience, even if human cognition turns out to be fully computational, the case for human dignity, for the irreducible value of human relationships, for the moral seriousness of forming young people, would still stand. Those arguments rest on commitments about how we should live and what we owe each other, not on contested empirical claims about machines. The encyclical would be stronger, not weaker, if it bracketed the metaphysics and stood on the ethics.

In short: I think Leo XIV is right that AI is not a person and should not be treated as one. I think he is also wrong to claim that we know, with certainty, why. If we are going to talk to our students about these systems — and we are — we owe them both the practical wisdom and the epistemic humility. Treat AI as a tool, not a friend. But also: we are genuinely uncertain about what these systems are, and that uncertainty is part of what makes this moment so demanding. Pretending we have certainty when we do not is its own form of dehumanization, because it denies young people the chance to think honestly about one of the deepest questions of their lifetime.

Subscribe now

3. The essential responsibility of education, and especially of school leaders

The encyclical does not leave educators ambiguous about their role. It names them directly, repeatedly, and with a weight most school leaders have not yet allowed themselves to feel.

In paragraph 13, Leo XIV walks through the Nehemiah image that opens the document — the rebuilding of Jerusalem’s walls — and makes its meaning unmistakable.

“Building a world in which everyone can flourish requires shared responsibility and courage. No one can single-handedly bear the weight of the challenges the world is facing, just as no one is so weak that they cannot play their part... All are given their own section of the wall: scientists and researchers, entrepreneurs and workers, educators and legislators, civil society, popular movements and faith communities.”

Each given their own section of the wall. This is the encyclical’s structural answer to the question that haunts every conversation about AI. Who is responsible? The answer is: everyone is responsible, but no one is responsible for everything. The scientist is responsible for the science. The entrepreneur is responsible for what gets built. The legislator is responsible for the framework. The worker is responsible for the labor. The educator is responsible for forming the people who will fill all those other roles.

Notice who Leo XIV puts on this list. Educators and legislators. He places them in the same clause, with the same conjunction, as if to insist that they carry comparable weight. This is not how the AI policy conversation is usually structured. The dominant conversation treats educators as objects of policy — recipients of guidance, implementers of frameworks, the people who have to figure out how to apply rules made elsewhere. The encyclical treats educators as agents.

You have a section of the wall. Not a memo about the wall. Not a workshop on the wall. A section.

Subscribe now

What our section looks like

Leo XIV is not vague about this. Thirty-four paragraphs after he places educators on the wall, he speaks to academic institutions directly. “I would like to encourage academic institutions and universities to give fresh impetus to these principles, and to apply them in a way that will be relevant and effective in addressing the digital revolution” (¶47).

Read what he is asking for and what he is not. He is not asking schools to adopt a particular AI policy. He is not asking universities to ban or to embrace the tools. He is asking institutions of learning to take the principles — human dignity, the common good, the universal destination of goods, subsidiarity, solidarity, social justice — and give them fresh impetus in the context of the digital revolution. Make them live again. Make them speak to what is actually happening. Translate them into the language of this moment so that students, faculty, and the broader public can recognize them as the criteria by which we judge what we are building.

This is a much more demanding charge than “write an AI policy.” A policy can be drafted in a weekend. Giving fresh impetus to a set of moral principles in the context of a civilizational transformation is the work of a generation.

So what is the educator’s section of the wall? I have been turning this over for the better part of two years, and I think it comes down to this.

We are responsible for the formation of the people who will live inside this technology for the rest of their lives. Not the technology itself — we are not building the models. Not the policy — most of us will never write legislation. Not the economics — we do not set the price of compute. Our section is human beings. The young ones, specifically. The ones who are forming, right now, the habits of mind and heart that will determine whether they bring openness, listening, judgment, courage, and solidarity to the decisions ahead of them — or whether they bring something thinner and more pliable.

This is harder than it sounds, because the technology is actively working against it. Every system the students use is being optimized to reduce friction. Every interface is being smoothed. Every cognitive task that used to require effort is being offered as a service. The implicit message of the consumer AI experience is that thinking, writing, deciding, struggling — these are inefficiencies to be eliminated. Our section of the wall is the place where we insist, against that message, that some kinds of friction are not bugs but features of becoming human. That writing the bad first draft is part of how the mind learns to think. That sitting with the hard question, rather than auto-completing past it, is what produces judgment. That arguing with a person who disagrees with you — really arguing, in real time, with stakes — is what builds the capacity for democratic citizenship that no chatbot conversation can simulate.

Subscribe now

Informed users as governance infrastructure

There is one more thing the encyclical makes explicit, and I want to flag it because it elevates what educators do from a soft adjacent concern to a constitutive part of how AI gets governed at all. In paragraph 106, Leo XIV is talking about what is required for AI to be governed in a way that respects human dignity — and he names four things, not three. “It is not enough to invoke ethics in the abstract; robust legal frameworks, independent oversight, informed users and a political system that does not abdicate its responsibility are required.”

Informed users. Read that as a category in its own right, sitting beside laws and regulators. The Pope is saying that even with the best legal frameworks and the most rigorous oversight bodies, an AI ecosystem populated by uninformed users will not be a humane one. The reason is structural. Laws lag. Regulators are outmatched. Oversight is partial. The frontier of what gets deployed and used moves faster than any of those mechanisms can catch up to. Which means that in the everyday encounter between a human being and an AI system, the only governance mechanism actually present in the room is the human being — and whether they understand what the system is, what it is doing, what it is optimizing for, and what they should and should not trust it to do.

That capacity does not arrive on its own. It is built. By teachers, in classrooms, over years. Informed users is what an education system produces when it is doing its job in this moment. The encyclical is, in effect, telling us that informed users are a piece of the governance infrastructure of AI, and that the institution responsible for producing them is the school. Not the only institution — families matter, communities matter, faith communities matter — but the institution with the explicit mandate, the dedicated time, and the trained professionals.

The legal frameworks need lawyers. The oversight needs auditors and engineers. The political system needs leaders who will not abdicate. And the informed users need educators. Without us, the rest does not work.

Subscribe now

The plural dimensions of formation

For most of my career in education, the dominant trend in American schooling has been to narrow what schools measure and reward toward what is testable, which in practice means toward what is cognitive. Reading scores. Math scores. AP scores. SAT scores. The metrics that capture intelligence — and a fairly thin slice of intelligence at that, the parts that are easy to assess on paper — have become the metrics by which we evaluate students, teachers, schools, and districts. Other things schools used to take seriously — the formation of character, the building of relationships, the cultivation of will, the development of emotional life, the practice of commitment, the slow work of moral imagination — have receded.

Paragraph 113 of the encyclical names this drift in language I have not seen elsewhere. Leo XIV writes: “elevating any single dimension of human existence to an absolute is always a mistake. Indeed, disorder does not arise only from scarcity; even unchecked growth can give rise to impoverishment. In an ecosystem, balance is disrupted when one species expands at the expense of others; in human life, something similar occurs when one faculty claims to be the measure of everything. Thus, intelligence, when absolutized, overshadows other essential dimensions of life, such as affection, the will, commitment, and relationships.”

Read that with a school in mind. Disorder does not arise only from scarcity; even unchecked growth can give rise to impoverishment. A school that has expanded its emphasis on the measurable cognitive outputs of education at the expense of everything else is not necessarily a failing school by the metrics it has chosen. It might be a thriving school by those metrics. Test scores up. AP enrollment up. College placement up. But the encyclical is saying that the metric itself is the problem.

Subscribe now

This was already a problem before AI. AI makes it acute, and in a particular way. Among all the dimensions of what a human being is — affection, the will, commitment, relationships, moral imagination, the capacity for sustained attention, the practice of love, the cultivation of taste, the formation of judgment — the one dimension AI can most plausibly simulate is the cognitive. The thing schools have been over-investing in for decades — narrow cognitive output — is the thing AI is most likely to substitute for.

The schools that have continued to take seriously the other dimensions Leo XIV names — affection, will, commitment, relationships — turn out to have been investing all along in exactly the human capacities that AI cannot substitute for. The character formation many of us thought of as the soft side of school is now, possibly, the hard side.

A school that takes paragraph 113 seriously stops treating relational, ethical, and characterological formation as the part of the day that has to fight for time against the academic priorities. It treats them as the academic priorities — not because the cognitive does not matter, but because the cognitive without the rest is exactly the impoverishment Leo XIV is naming.

Subscribe now

The four-part ecology of communication

The encyclical then gets specific about what this looks like in practice. In paragraph 137, Leo XIV writes: “Our first task is neither to demonize nor idolize technological tools, but to utilize them on the basis of a fundamental principle, namely that truth is a common good and not the property of those with power or influence. We must therefore promote an ecology of communication. On the level of public policy, this entails establishing norms so that the decision-making behind content selection and its development becomes more transparent and protects personal data. Regarding social and cultural aspects, this requires a strengthening of intermediary organizations, serious journalism and forums for debate, where reasoned argumentation and verification carry greater weight than immediate reaction. For families and schools, there is a growing need for new educational awareness and for formation concerning the proper and critical use of digital tools, AI and online commercial and financial platforms. In universities, the principal challenge lies in the integration of knowledge, cultivating both the capacity to connect and synthesize knowledge in order to grasp complexity, and the skills necessary to verify facts.”

The Pope is sketching a four-part ecology: public policy, civil society, schools and families, universities. Each one has a specific assignment. None of them can do the work alone.

I want to flag one phrase in particular. Forums for debate. The Pope is naming, as one of the three pillars of the social and cultural response to the AI moment, the kind of structured argumentative practice that competitive debate has been refining for over a century. Reasoned argumentation and verification carrying greater weight than immediate reaction is, in fact, the entire pedagogical theory of debate education in a single phrase. The reaction economy of social media is what debate, at its best, is constitutively organized against.

For universities, the encyclical’s charge is the sharpest of the four. The principal challenge — primary, central — is the integration of knowledge. Not its production. Not its dissemination. Not its credentialing. Its integration. The capacity to connect, synthesize, and grasp complexity across domains. This is the structural indictment of the modern research university, which has spent the last sixty years organizing itself toward the opposite of what the encyclical is asking for. AI can perform a great deal of the narrow cognitive work each discipline has trained itself to value. What it cannot do — at least not yet, and possibly not ever in the senses that matter — is the integrative work of bringing knowledge from multiple domains to bear on a complex situation, holding the tensions honestly, and producing judgment. That work is the work humans will continue to need to do. It is also the work that the universities are least well-organized to teach.

Subscribe now

Education is decisive, and we are unprepared

Paragraphs 139 through 142 are where the encyclical drives the educational dimension home with a precision most readers will not be prepared for.

Paragraph 139 opens with a sentence that should stop any educator who reads it. “In an era when truth is often distorted in order to serve particular interests and communication strategies, the field of education assumes decisive importance. Yet rapid technological transformations reveal just how unprepared we are on the educational level.”

Decisive importance. Then, in the same breath, we are unprepared. That juxtaposition is the situation in one sentence. The work that education has to do in the AI moment is more consequential than at any point in recent memory, and the institutions that are supposed to do that work are not yet doing it.

Paragraph 140 contains the sharpest pedagogical claim in the entire document. “Education, by contrast, is a long journey requiring patience, and therefore needs time for development and for engagement with reality beyond appearances. This is a fundamental issue because every technology shapes those who use it. Educating people about the use of AI, then, involves teaching them to decide when and for what purpose it ought not to be used.”

The dominant frame is how to use AI well. The encyclical is saying that the more important question is when to refuse to use it at all. A graduate who has been taught only how to use these tools is half-formed. A graduate who has been taught when not to use them — and why — is the kind of person who can remain human in a world full of machines that want to think for them.

Then Leo XIV reaches for an unexpected source. “As Plato wrote, the deepest and most important things are learned only after much time and effort, by engaging in discussion with others, ‘striking upon’ ideas and experiences together like flint until the spark of understanding is kindled within us.”

The image is striking and worth dwelling on. Flint striking flint until the spark of understanding is kindled. This is not the model of education that most of the technology industry is selling us. The technology industry is selling instantaneous answers, frictionless interfaces, personalized tutors that adapt to the student in real time, summaries that compress hours of reading into minutes. The encyclical is reaching back twenty-four centuries to Plato to remind us that the deepest and most important things do not work that way.

The paragraph closes with the line that I think should be quoted on the wall of every school. “We must learn, then, how to exercise restraint in the use of AI and to protect our young people from the promise of the perfect machine, from that subtle temptation which renders human thought seemingly superfluous precisely when it is most needed.”

The promise of the perfect machine. Paragraph 141 then catalogs what is happening to young people inside the current digital environment, in language unusually direct for an encyclical — sleep destroyed, attention destroyed, emotional regulation destroyed, the pornography crisis, the grooming crisis, the deepfake crisis, the eating disorder crisis, the suicide crisis. AI is named explicitly as an aggravating factor. The point is structural. The same companies, the same business models, the same engagement-maximizing logic that produced the smartphone-era harms are the ones now building the AI systems.

Paragraph 142 is the constructive turn. “It is difficult for parents by themselves to resist the influence of business models that monetize attention and time. Therefore, it is essential to form an alliance among policy-makers, educational institutions and families that is capable of concretely supporting adults in this task.” The alliance the Pope is proposing is not metaphorical. It is concrete: legislators, schools, and families, working together, against business models that have proven they will not voluntarily reform.

Subscribe now

The special responsibility of school leaders

There is one move the encyclical makes that I want to spend some time on, because it identifies the people in education who carry a specific weight in this moment.

In paragraph 111, Leo XIV addresses AI developers directly. “I wish to address a special appeal to those who develop artificial intelligence. In one sense, technological innovation can represent human participation in the divine act of creation. Developers, therefore, bear a particular ethical and spiritual responsibility, for every design choice reflects a vision of humanity. Just as the creator of an artistic or literary work must consider the values it conveys, so developers are called to embed values in their projects with due seriousness: with transparency, responsibility toward affected communities and careful attention to ensuring that what is being cultivated is a genuine good.”

The most important word in that passage is cultivated. The Pope is not saying that AI is built like a building. He is saying it is cultivated — like a garden, like a child, like a tradition.

Now read that paragraph again and replace “AI developers” with “heads of school.” Or superintendents, or college presidents, or department chairs. The analogy is exact.

Heads of school cultivate. They do not build. A school is not a thing you assemble from a kit; it is a culture you tend, year after year, through countless decisions about what gets prioritized, what gets measured, what gets celebrated, what gets quietly dropped. Every choice the head makes — what to put in the budget, who to hire, what to assess, what to defend in front of the board, which parent to push back on, which faculty member to back, which technology to adopt, which to refuse — reflects a vision of what the school exists to do and what kind of human being the school is trying to form. Just as the developer’s design choices embed a vision of humanity in the system, the head’s institutional choices embed a vision of humanity in the school.

I want to sit with the weight of this for a moment, because I do not think most heads of school have allowed themselves to feel it.

A head of school of even a modestly sized school is responsible, over the course of a tenure, for the formation of thousands of young people. Not their grades. Not their college placements. Their formation — the habits of mind, the dispositions of heart, the capacities for judgment, the relationships with their own attention and their own thinking, the moral imagination they will carry into their work and their citizenship and their families for the next sixty or seventy years.

In this moment, a head of school will be leading thousands of souls through what is almost certainly the most significant technological transition in human history. Not one of the most significant — the most. Industrial mechanization reshaped what people did with their bodies. The internet reshaped what people did with information. Artificial intelligence is reshaping what people do with their minds — what thinking even is, what judgment even consists of, what counts as human work and what does not. The students sitting in school today are the first generation that will spend their entire adult lives inside this transition. The shape they take during their school years will determine whether they pass through it as agents or as objects, whether they emerge from it more deeply human or thinner, whether they recognize the moral choices the technology embeds or accept those choices as the water they swim in.

That is the work in front of a head of school right now. The school had them for thirteen years, six hours a day, during the formative window. Whatever the school decided to cultivate in them is what most of them will become.

This is the responsibility the encyclical is asking heads of school to feel. Not the responsibility of a manager. Not the responsibility of an administrator. The responsibility of someone who has been entrusted with thousands of souls during the most consequential transition their species has yet faced.

Subscribe now

The operational specifics

The encyclical itself, in paragraph 145, supplies the operational direction:

“The second major challenge is pedagogical. Many educational systems struggle to keep pace with change and to support the integral development of students. The advance of information technologies and AI is rapidly rendering curricula obsolete that were designed for a different era. Meanwhile, the organization of schools, physical spaces, evaluation methods and the role of teachers themselves must be rethought in order to promote an authentically integral education that addresses every dimension of the person. It is necessary to support the ongoing formation of teachers throughout their professional lives, so that they can engage positively with new technologies, helping students to use them responsibly, critically and creatively, rather than passively succumbing to their influence.”

Read what the encyclical is putting on the table. Curricula being rendered obsolete. The organization of schools requiring rethinking. Physical spaces requiring rethinking. Evaluation methods requiring rethinking. The role of teachers requiring rethinking. The ongoing formation of teachers throughout their professional lives. Six things. The encyclical is not asking heads of school to choose which ones to take up. It is naming all six as the work the moment requires. A head of school who is doing one of them well and ignoring the others is still leaving most of the work undone.

The student passing through the school is not a collection of test scores. The student is a whole human being — cognitive and affective, intellectual and relational, embodied and moral — and the school either takes responsibility for the formation of all of those dimensions or it does not. Authentically integral education that addresses every dimension of the person. That is the standard the encyclical sets. The heads of school who are willing to meet it are the ones who will, when the historical work of this period gets written, be remembered as the ones who actually rose to the moment.

The rest of them will have been administrators.

Subscribe now

No one is without responsibility

I can hear the objection forming in some readers by this point, because I have felt it myself in writing it. The problems are too big. The institutions are too entrenched. I am too small. The decisions I make in my classroom or my school or my home will be drowned out by the scale of what is happening at the labs, in the markets, in the platforms. What is the point of taking on this much weight if the weight I can move is so much smaller than the weight that is moving everything else?

The encyclical anticipates this objection and refuses it. In paragraph 212, Leo XIV writes:

“At this point, however, a subtle temptation may emerge, namely the thought that the problems are too big and we are too small, and that our choices, therefore, cannot make a difference. This is a polite form of resignation, often disguised as realism. Certainly, not everyone has the same power to make a difference. There are those who govern, make investment decisions, lead institutions, conduct research, educate, produce or provide information, and then there are those who only seem to live their daily lives. Yet, no one is without responsibility.”

A polite form of resignation, often disguised as realism. The Pope is naming the move precisely. The thought that the work is too big for the worker is not a humble recognition of limitations. It is, in the encyclical’s framing, a form of giving up that has learned to dress itself as wisdom.

Notice who Leo XIV names in the list of people whose decisions matter. Those who govern. Those who make investment decisions. Those who lead institutions. Those who conduct research. Those who educate. Those who produce or provide information. Read that list carefully. Those who educate sits in the middle of it, between researchers and information providers, alongside investors and governors. The encyclical is not placing educators at the periphery of consequence. It is placing them in the central list of people who have the power to make a difference.

The work is not too big for you to carry your part of it. The institutions are not too entrenched for your choices inside them to matter. The platforms are not too powerful for the formation you provide to be the structural counterweight that the encyclical is asking schools to be. You are not too small. The encyclical does not allow you to be too small.

No one is without responsibility.

Subscribe now

4. The risks the encyclical names

One of the most striking things about Magnifica Humanitas is that it does not approach AI from a position of detachment. The encyclical engages, point by point, with the specific worries that ordinary people, journalists, and researchers have been raising for years. Seeing them collected in one place reveals how comprehensive the encyclical’s coverage really is.

Loneliness and parasocial relationships with AI

The encyclical takes seriously the now-documented phenomenon of users forming emotional attachments to AI systems. “The artificial imitation of positive human communication — words of advice, empathy, friendship and even love — can be engaging and at times genuinely helpful. However, for less discerning users, it can also be misleading, creating the illusion of a relationship with a real personal subject. When words are simulated, they do not build genuine relationships, but only their appearance” (¶100). The risk that what looks like connection becomes a substitute for the harder work of building real human bonds — and that the substitution is most dangerous for people who are already alone.

Environmental cost

Leo XIV is explicit about the environmental footprint of AI infrastructure. “Current AI systems require enormous amounts of energy and water, significantly influencing carbon dioxide emissions, and place heavy demands on natural resources” (¶101). This is the part of the AI discourse that techno-optimists tend to wave away. The encyclical refuses to.

Subscribe now

Automated decisions and the exclusion of the vulnerable

Two paragraphs name the algorithmic decision-making problem with unusual precision. “Important and sensitive decisions — concerning employment, credit, access to public services or even a person’s reputation — risk being fully delegated to automated systems that do not know ‘compassion, mercy, forgiveness, and above all, the hope that people are able to change,’ and can therefore give rise to new forms of exclusion” (¶102). Paragraph 103 sharpens it: when an algorithm gets to decide who is worthy without anyone bearing responsibility for that judgment, the exclusion of the vulnerable becomes cloaked in a veneer of neutrality and objectivity, against which it becomes difficult to raise objections. The cruelty does not look like cruelty. It looks like math.

Privacy and surveillance

Privacy is named directly in paragraph 102 as a category of clearly harmful use. The deeper treatment is in paragraph 171, where the encyclical names the surveillance dimension as a structural threat to freedom itself. The mass collection of behavioral data — movements, purchases, relationships, preferences — produces the power to profile, predict and influence behavior, often without individuals being fully aware of it.

Deepfakes and the manipulation of reality

The encyclical names the deepfake problem directly. “The ability to manipulate content, images and videos exposes people to biased or misleading perspectives” (¶132). And it returns to the question in the war context (¶¶190-191) where AI-enabled manipulation of images, video, and information becomes a tool of conflict.

Unemployment, generally and for the young

The encyclical’s treatment of unemployment is one of the most extended in the document. Paragraphs 151 through 156 walk through what Leo XIV calls a true social calamity — the prospect of significant and rapid contraction in available jobs as AI replaces what humans used to do. The Pope explicitly names the fourth industrial revolution and warns of new forms of job insecurity and inequality, characterized by outsized remuneration for a highly specialized minority alongside declining wages for a large portion of the workforce. Then in paragraphs 165 through 167, he turns to the specific impact on young people, arguing that job insecurity at the start of working life is particularly devastating because work is the context in which identity is formed, vocation is discerned, and adulthood begins.

Subscribe now

Slavery and worker abuse in the AI supply chain

This is the part of the encyclical that will shock readers who have not been paying attention to the AI labor question. Paragraphs 173 and 174 name what is happening in the production of AI systems and the devices that run them. The data labelers, the content moderators reviewing traumatizing material, the rare earth miners (often children) producing the materials in modern electronics — Leo XIV calls these new forms of slavery directly linked to the digital economy and refuses to let them remain invisible. “Nothing in the world of AI is immaterial or magical. Every seemingly immediate and flawless response is the result of a long chain of mediation, involving vast networks of natural resources, energy infrastructure and, above all, people” (¶173).

Data colonialism

Paragraph 178 introduces a term most readers will not have encountered in religious documents before: data colonialism. The encyclical argues that AI has produced a new form of extraction in which entire populations have their personal lives transformed into exploitable information, with the wealth produced flowing back to the centers of technological power.

War and cyber war

Paragraphs 183 and 197 are the encyclical’s treatment of AI in armed conflict. Leo XIV refuses the framing that algorithmic targeting and autonomous weapons are technical questions to be resolved by engineers. They are moral questions about whether human beings will continue to bear responsibility for the decision to take life. Paragraph 225 extends this concern to the cyber domain — the AI-enabled disruption of infrastructure, communications, financial systems, and democratic processes.

AI replacing moral judgment

Perhaps the deepest of the encyclical’s warnings. Paragraph 198 addresses the temptation to delegate moral judgment itself to AI systems — to let the algorithm decide what is right and wrong, who deserves what, how scarce resources should be allocated, who should be punished, who should be helped. The encyclical insists that moral judgment is a constitutively human act, that no system can bear the responsibility of judgment because no system can bear the consequences of getting it wrong.

Subscribe now

What the catalog adds up to

The encyclical is doing something more than listing concerns. It is demonstrating that the worries ordinary people have about AI are not paranoid, not Luddite, not nostalgic, not anti-progress. They are correctly identified, and they have moral weight that the encyclical is willing to put behind them.

This matters because the dominant move in AI industry communication is to treat critics as the people who do not understand the technology. They are afraid because they are uninformed. They will come around once they see the benefits. The concerns are overblown. The encyclical, working from outside the industry’s framing, refuses this dismissal.

A reader who finishes the encyclical and finds their own AI-related worries reflected back at them, named with theological weight, and treated as worthy of papal attention is having an unusual experience. Most institutional discourse about AI has trained that reader to feel that their concerns are quaint. The encyclical treats them as the central material of a major intellectual document.

The Pope, in other words, is on the side of the people who have been telling us all along that something is wrong here.

Subscribe now

5. The criteria — what the encyclical asks us to evaluate AI by

I have spent the bulk of this post on the institutional and educational dimensions of the encyclical. But Magnifica Humanitas is structured around a much larger inventory of principles, and I do not want to leave them unmentioned. The encyclical is doing something most contemporary AI commentary refuses to do, which is to insist that AI must be evaluated against a complete moral framework, not against whichever single value is convenient for the speaker.

The whole document, in some sense, comes down to a test that Leo XIV poses in paragraph 85. Technological innovations, he writes, including artificial intelligence, are not neutral. They must be evaluated by asking a crucial question: “Do they truly help individuals and peoples to become more humane and fraternal, while respecting our common home and future generations?”

That is the question. Every principle below is, in effect, a different angle on the same examination.

Subscribe now

Dignity and the rejection of “lower-value human capital”

The foundational principle is human dignity, and the encyclical insists on it in language that will not bend. “The value of persons does not depend on what they achieve or produce. There are rights that apply to everyone simply by virtue of being human, and no human power can legitimately deny or arbitrarily limit them” (¶51). And again: “persons end up being reduced to a means of achieving results, a resource to be used and exploited, and are no longer recognized as a proper end in themselves who should never be instrumentalized” (¶51).

This is the principle that gets violated in plain sight on a regular basis in AI-driven corporate communication. Last week, Standard Chartered CEO Bill Winters, explaining the bank’s plan to eliminate 7,800 jobs through AI, told investors: “It’s not cost cutting. It’s replacing, in some cases, lower-value human capital with the financial capital and the investment capital we’re putting in.” The phrase produced enough backlash that Winters apologized for it within days, but the apology missed the point. The slip was not a word choice. It was a worldview slipping into view. Lower-value human capital is what the encyclical’s paragraph 51 names: persons reduced to a means of achieving results.

Leo XIV’s response, three paragraphs later, is unsparing: “Every human person possesses an infinite dignity, inalienably grounded in his or her very being, which prevails in and beyond every circumstance, state, or situation the person may ever encounter” (¶53). Infinite dignity is not a phrase that fits on an investor slide. It is, however, the phrase the encyclical insists belongs at the center of the conversation.

Subscribe now

Work as constitutive of dignity

If dignity is foundational, work is the place dignity becomes concrete. From paragraph 30, quoting Leo XIII: persons have “a fundamental value that takes precedence over capital and profit.” From paragraph 37, on John Paul II’s Laborem Exercens: “Work is not considered simply as a problem to be dealt with or a means of generating income, but a fundamental good for the person, a principle of economic activity and the key to the entire societal question.”

This frames the entire AI-and-labor conversation differently than most policy discussions do. The standard framing treats jobs as the dependent variable. The encyclical insists that work itself, not the wage attached to it, is a constitutive part of being human. “Through work, human beings bring their freedom, creativity and capacity for cooperation into play, contributing to the cultural and moral elevation of society” (¶37). When AI eliminates a job, it does not just eliminate a paycheck. It eliminates one of the structures through which a human being expresses freedom and creativity.

Subsidiarity in the age of platforms

Paragraph 71 is, to my mind, the most directly relevant single paragraph in the entire encyclical for AI governance. “The principle of subsidiarity applies especially in the context of the digital revolution. Here, the highest level is not the State, but rather major economic and technological actors that exercise de facto power over the conditions of everyday life... The principle of subsidiarity requires that such processes not be imposed from above in an opaque and unilateral manner, but instead be directed toward the common good with transparency, accountability and meaningful forms of participation (including independent checks, transparency regarding algorithms, equitable access to data and avenues for recourse).”

The Pope is saying that the traditional architecture of subsidiarity — the State at the top, intermediate organizations below, individuals and families at the base — has been displaced. The new top of the pyramid is private. And the principle of subsidiarity, properly applied, demands that this new top tier operate with the same transparency, accountability, and openness to participation that we would demand of any government wielding comparable power.

Subscribe now

Solidarity, social justice, and the universal destination of goods

The encyclical names solidarity as both a principle and a virtue (¶75) — “a ‘firm and persevering determination’ to strive for the common good, with particular attention to those most in need.” Applied to AI, this is paragraph 76: “Solidarity demands that decisions regarding data, algorithms, platforms and artificial intelligence take into account not only the immediate benefit for a few, but also the impact on all peoples and on future generations.”

Social justice in the AI moment, paragraph 80: “Justice demands that we prevent the emergence of new forms of exclusion and deprivation of freedoms: individuals and peoples hindered or denied access to basic technologies, communities exposed to invasive surveillance and social groups penalized by opaque algorithms that perpetuate prejudice and discrimination.”

The universal destination of goods, paragraph 67, extends old principles to new property: “Today, among the goods that are universally intended for everyone, we must also include new forms of property, such as patents, algorithms, digital platforms, technological infrastructure and data. In a context where the wealth of nations depends increasingly on knowledge and technology, when these goods remain concentrated in the hands of a few, without adequate forms of sharing and access, a new imbalance is created that contradicts the universal destination of goods.”

And in paragraph 108, the proposal that I think is the most economically interesting move in the entire document: “Ownership of data cannot be left solely in private hands but must be appropriately regulated. Data is the product of many contributors and should not be treated as something to be sold off or entrusted to a select few. It is necessary to think creatively in order to manage data as a common or shared good, in a spirit of participation.” Data is the product of many contributors and ought to be managed as a common or shared good. This lands in direct opposition to the operating assumption of essentially every major AI company, which is that the data on their servers is theirs to monetize.

Subscribe now

Access as the operational test

In paragraph 109, Leo XIV does something I have not seen in any other piece of writing on this subject. He takes five separate principles of social doctrine, walks them through one by one, and applies each of them specifically to the question of access.

“In a world where data, computational resources and regulatory influence remain in the hands of a few, to speak of the common good means exposing this new form of epistemic, economic and political asymmetry and naming the new monopolies of AI. To speak of the universal destination of goods means finding ways of ensuring universal access to both technologies and the education needed to use them. To speak of subsidiarity calls for protecting the ability of communities to make choices and corrections, rather than confining their role to mere oversight after the standards have been set elsewhere. To speak of solidarity obliges us to recognize the hidden, often exploited workers, who sustain algorithmic systems. To speak of justice requires questioning the global distribution of power that decides who in fact can train these models and who is merely subjected to them.”

Five principles. Each one applied to AI. Each one centered on the same question: who has access, and who does not. The Pope is treating access as the place where the entire framework of social doctrine either lands in practice or fails to. If access is not real, then the common good is a slogan, the universal destination of goods is a phrase, subsidiarity is a fig leaf, solidarity is performance, and justice is decoration.

I want to apply this directly to something happening right now in American schools that almost no one is framing in these terms. During the pandemic, school districts distributed millions of laptops and tablets to students. Nearly nine in ten public schools provide a device for each middle and high school student, and more than four in five elementary schools do the same. For students in low-income households, these devices represented something genuinely new. That window is now closing.

In Wake County, North Carolina, the school system serves 160,000 students. The superintendent told the board in early 2026 that “One-to-one is something that a district of our size just cannot afford.” In Fresno Unified, the change has already happened. “No longer will every student in the Fresno Unified School District have a laptop for use at home,” after six years of universal access.

I understand the budget pressures. They are real. But none of this changes what the encyclical is naming in paragraph 109. The question is not whether the budget pressures are real. The question is whether the response to those pressures is consistent with the principles the encyclical insists must structure these decisions. The decision to take the device back from the students who have the least access at home widens the asymmetry. It is the institutional act of letting the inequality grow. The encyclical’s framework is asking us to notice that. Most of the current discourse about K-12 device programs is not.

Subscribe now

Democracy, dialogue, and the common good

The encyclical is clear that democracy itself is at stake. Paragraph 39, summarizing John Paul II in Centesimus Annus: “the Church values democracy insofar as it guarantees the effective participation of citizens, enables them to elect and peacefully replace their leaders and prevents power from being monopolized by small elite groups motivated by particular or ideological interests.” Paragraph 63 names the danger directly: “When politics abandons a long-term perspective and reduces itself to short-term calculations or sterile polarizations, then the language of the common good loses credibility.”

And the encyclical insists, against the dominant register of inevitability, that humanity has built cooperative institutions before. Paragraph 123: “History does not appear solely as a record of human violence, but also as evidence that humanity is capable of creating institutions that protect our shared life. Over the past two centuries, this can be seen in several emblematic achievements: the founding of the International Committee of the Red Cross (1863), whose operational neutrality ensures compassionate care for all; the long process that led to the abolition of slavery, which represented not only a legal shift but a transformation of conscience; the establishment of the United Nations (1945) and the Universal Declaration of Human Rights (1948), which articulated a shared language for affirming, at least as a common ideal, the universality of human dignity; and the 1951 Refugee Convention, which recognizes the duty to protect those fleeing persecution and danger.”

The Red Cross was not inevitable. The abolition of slavery was not inevitable. The United Nations was not inevitable. The Refugee Convention was not inevitable. Each of these institutions represents a moment when humanity, faced with what looked like an immovable structural problem, built something anyway.

The encyclical is not naive. It does not promise that the AI moment will resolve into a humane outcome. It promises that the possibility of a humane outcome has not been foreclosed.

Integral human development

The synthesizing principle, paragraph 82, returns to Paul VI: development is authentic only if it is “integral,” “foster[ing] the development of each man and of the whole man.” Paragraph 83 spells out what that means: “Development is truly human when it places people at the center instead of the accumulation of wealth, and when it concerns peoples as well as individuals.” Paragraph 84 ties development to ecology and intergenerational justice: “true progress is not what increases the wellbeing of some by degrading ecosystems, shifting costs onto the most disadvantaged communities, or compromising the living conditions of those who will follow us.”

This is the encyclical’s most demanding standard, because it forecloses the easy out. You cannot defend an AI buildout by pointing to the abundance it produces in one quadrant if it degrades ecosystems in another, concentrates wealth in a third, and forecloses options for the next generation in a fourth. Integral means the whole picture has to add up.

Subscribe now

Human rights

I will close this survey where the encyclical does, with rights. Paragraph 54: “the movement toward the identification and proclamation of human rights is one of the most significant attempts to respond effectively to the inescapable demands of human dignity.” Paragraph 56 names the modern danger directly: “the protection of human rights has been exposed to two particularly serious dangers. The first is that these rights are declared in a purely formal sense, while technological progress continues alongside covert or overt violations of human dignity.”

That last sentence is the encyclical’s prophecy, and it is already coming true. We declare rights in the abstract while the systems being built around us violate them in practice. The right to privacy declared formally; surveillance capitalism advancing as if the right did not exist. The right to a fair hearing declared formally; algorithmic decision-making penetrating credit, employment, housing, and criminal justice with negligible review.

The integrating test

The principles are many. The test is one. Do they truly help individuals and peoples to become more humane and fraternal, while respecting our common home and future generations?

Run any current AI deployment through that question, honestly, and notice how many of them struggle to give an unambiguous yes. That is not an argument for stopping. It is an argument for the seriousness the encyclical models — for refusing to let “abundance” or “innovation” or “competitiveness” stand in for the harder question of what we are actually building, for whom, and at what cost.

6. What to make of all this

I have spent a long time inside this document, and I want to close by saying what I think it is for.

Magnifica Humanitas is not a policy paper. It does not tell us which AI systems to ban, which to permit, which to regulate, and which to subsidize. It does not tell us what the right tax rate on AI companies is, or how to write an EU AI Act, or whether students should be allowed to use ChatGPT for their college essays. The document is not in the business of producing those answers, and reading it for those answers is reading it wrong.

What the encyclical does do is give us a framework. A vocabulary. A set of questions we can carry into the rooms where the actual decisions get made — the board meeting, the faculty meeting, the legislative session, the dinner table, the conversation with our own children. The framework is durable. The principles have been refined for more than a century of social teaching. They have already withstood the industrial revolution, two world wars, the rise and fall of competing ideologies, the emergence of globalization, and the climate crisis. They are not the only framework available, and you can argue with them, but they are a framework — coherent, tested, and capable of orienting people who do not otherwise know how to think about what is happening.

This matters because the dominant feature of the current AI moment, more than any technical capability or commercial development, is disorientation. Most people I talk to — educators, parents, students, professionals — feel that something significant is happening but do not know how to assess it. The technology industry’s framing is exciting but unconvincing. The doomer framing is alarming but unactionable. The policy framing is bureaucratic but slow. The popular framing is fragmented across a hundred think pieces that contradict each other. The result is that thoughtful people are largely watching, hoping for the best, and not acting because they do not know how to act.

The encyclical is, among other things, an act of orientation. It says: here are the questions to ask, here are the principles to apply, here are the people whose responsibility this is, here is the work in front of us. You do not have to share the theology to take the orientation seriously. The framework holds even if you bracket the Catholic frame — it is one of the few places in contemporary discourse where the moral structure of what is happening has been worked through in detail.

For educators in particular, I think the encyclical does three things that nothing else currently available does.

It tells us that what we do matters. That the work of forming young human beings is one of the central tasks of this civilizational moment, not a peripheral concern. That the institutions we lead are not downstream of the labs and the platforms — they are upstream of the citizens who will, in twenty years, be the engineers and executives and voters and parents whose decisions will determine what these systems become. Those who govern, those who make investment decisions, those who lead institutions, those who conduct research, those who educate, those who produce or provide information. We are on the list.

It tells us what the work actually requires. Not a curriculum unit on AI literacy. Not an acceptable use policy. The full reorganization of curriculum, school structure, physical space, assessment, the role of the teacher, and the ongoing formation of teachers — all of it directed at producing human beings whose interior freedom is intact, whose capacity for sustained attention has been cultivated, whose moral imagination has been formed, and who can recognize the manipulations of the platforms and refuse them. That work is much bigger than most schools are currently doing. The encyclical says it anyway.

And it tells us that we cannot wait for someone else to do the work for us. No one is without responsibility. The polite resignation that has become the dominant educator response — the technology is moving too fast, the institutions are too entrenched, we will figure it out eventually — is named in the document for what it is. A form of giving up that has learned to dress itself as realism.

The document is, finally, hopeful. Not naively. The encyclical does not promise that the AI moment will resolve well, and several of its sections — on warfare, on data colonialism, on the exploitation of children in the digital economy — are as grim as any contemporary writing on the subject. But it insists, against the inevitabilism that pervades both AI optimism and AI pessimism, that the future is not yet written. The Red Cross was not inevitable. The abolition of slavery was not inevitable. The United Nations was not inevitable. They happened because particular people, over decades, did the patient and uncertain work of building them. The same work is available to us now.

What the encyclical refuses to allow is the comfort of believing the work is too big for us. It is big. It is also distributable. Each of us has a section of the wall. The wall does not depend on any one of us alone. It does depend on each of us being willing to build the section we have been given.

For me, that section is the work of debate education — the patient, four-decade work of teaching young people to think carefully under pressure, to listen to people who disagree with them, to construct arguments grounded in evidence, to revise their positions when the reasoning requires it, and to remain in the room when the conversation gets hard. The encyclical tells me, more clearly than any document I have read in years, that this work is not adjacent to the AI moment. It is one of the practices the encyclical names by name — forums for debate, where reasoned argumentation and verification carry greater weight than immediate reaction — as essential to a humane response. That is my section.

What is yours?

The Pope is treating AI as a hinge of history. Magnifica Humanitas is the most serious document yet written on what that hinge means for human beings, what is at stake, what the principles are, and what the work consists of. Read it. Argue with it. Disagree where you must. But do not finish it without asking the question it ends on.

What is your section of the wall, and have you started building?

Subscribe now

AI Updates & AGI | Podcast Guests | AI “Literacy” |Instructional Redesign | Debate | Entrepreneurship | PBL | Societal Issues and Ethics | Unemployment, UBI, Social Contract | AI & Universities |

I spent some time today reorganizing and updating this Substack to make it easier for readers to engage with the material, which now includes 500 posts with over 1 million page views. Hopefully, the changes are obvious on the home page.

Thanks to everyone who has supported it, which is now almost 5,000 of you and 57 other Substacks.

Subscribe now

I started this Substack a little three years ago yesterday to think through how AI will disrupt education — and how we might best prepare students for a world where machines, increasingly embodied, match or surpass human intelligence in many domains (AGI/ASI). This is something like to happen in the next 3-4 years (or a bit sooner or later). (Hassabis)

My sense is that many of my readers are education reformers who want to move schools beyond a model built around preparing students to become “knowledge workers,” especially when that model still relies heavily on short-term recall and the production of knowledge as the main signs of learning. The most popular post on the site is MIT Study Shows Assigning Essays Triggers Student “Brain Rot,” and the 5th most popular, published only a week ago, is Institutionalized Education as Cognitive Offloading.

The Substack is a push toward the kind of teaching and learning John Dewey called for long ago: active, social, experiential, problem-centered, and connected to life beyond school. These ideas move away from standardized curricula that value everyone learning the same thing via direct instruction and then pit students against each other to see who can learn it the fastest and a little more accurately, even though AI can already offset any difference (Everyone is Above Average).

This call is not new. Educators have been making it for decades. What is new is the pressure. Generative AI can now produce much of the content schools have traditionally asked students to produce, often at a very high level. If the economic value of routine knowledge work is uncertain, then the educational model built around producing future knowledge workers is uncertain too.

Recent “traditional academic” manifestations of my approach include In Search of Deeper Learning by Harvard professors Mehta and Fine, and are reflected by the work of many others in this space, including Dr. Sabba Quidwai (see our book Humanity Amplified (available to paid subscribers of this Substack) and her Designing Schools: How Design Thinking Makes YOU Irreplaceable in the Age of AI, Dan Fitzpatrick, Jerry Crisci (perhaps the original “instructional redesigner,” with a large impact during his time at Scarsdale public schools), Phil Alcock, Dr. Alan Coverstone, and many others. Ultimately, I see AI, especially agentic AI, as a collaborator along a person’s educational journey, something also articulated in Humanity Amplified and in the work of Dr. Mairead Pratschke (paper, our podcast interview) and early work by Mollick & Mollick. It manifests itself in the afternoons of Alpha School and some of the curriculum at Qualia. There are many others along the continuum.

This is not in any way an “edtech” blog. AI as edtech can be powerful, and students can now easily build their own apps to support learning and especially studying, but before we focus on using AI as edtech in the current curriculum, we must first decide what the “curriculum” should be. If we just use AI to reinforce the current curriculum, we may achieve nothing, and we may even leave our students worse off.

There is certainly resistance to my way of thinking, and it is well-meaning. Ultimately, I think it stems from a belief that the way we’ve teaching and learning for the last 100-200 years is really awesome/the way to do it/likely the best way it can be done and that GAI simply allows students to offload the production of the artifact that was supposed to be the signal of what the learned, not only (further) distorting the signal but also undermining their own thinking and cognitive development. There has certainly been a lot of ink spilled on this, and even research done on it, but in many ways, it states the obvious — if the primary way educators get you to think and learn is by producing products, and the AI produces the product, you won’t learn the content and learn how to think. Duh, of course. We don’t need studies to prove that. The question is how to change the teaching and instruction not only so it is relevant in an AI World (Gates), but also so that humans will collaborate with each other and machines in ways that builds their intellectual muscle.

Subscribe now

What I’m calling for is not the (rapid) integration of AI into the current curriculum/system. That will produce many of the problems I just shared, but a redesign of the curriculum so AI can be integrated as a partner (peer, mentor, coach, teacher) to help students learn both content and skills that are relevant in a world of AI and other rapidly advancing technologies.

As we get closer to AGI, these technologies will continue to alter our economy, politics, social systems, and even who we understand ourselves to be as humans. To be frank, the idea that an educational system that was even starting to lose relevance in the past industrial era is going to be relevant in this one is almost absurd.

See posts on Instructional Redesign. Follow Patrick Dempsey on LinkedIn and pull the band-aid off.

Moreover, there is a false assumption that education is “ahistorical” — that regardless of the times we live in, we should teach and learn the same way. The reality is that education has always changed based on the socioeconomic and political system we live in. At least the first two of those are radically changing, and education will change with it.

Change, of course, is hard, even if AI is an arrival technology (MIT) and Agents are Already Inside our Schools. Even schools that want to change will struggle. As I outline in We Rebuilt the Military for AI. Schools Are Still Forming Committees, and Can Humans Launch a new course? Many wrongly believe the primary focus of AI in education is on what is the best chatbot to integrate rather than A Cognitive Layer, and they use normal decision-making structures to adapt to a world that is changing faster than those decision-making structures can respond to. And, the truth is that many don’t believe, or at least want to accept, that AI can do what it is already doing, let alone what it is becoming. It’s almost like being a parent who continues to educate and relate to their child as if she were a year old, rather than one who is 15 and will inevitably graduate from high school.

Schools are understandably struggling to keep up with each new iteration of AI. But the bigger challenge is not tracking every tool. It is accepting the direction of travel. We do not know the exact timeline, but we do know where this is headed: toward AI that is more capable, more conversational, more autonomous, more multimodal, and eventually more embodied. Students will increasingly live and work with systems that can reason, create, advise, tutor, simulate, decide, and act at or above human levels in many domains. Schools need to prepare students for that world, not merely for the next version of today’s chatbot.

Dr. Peter Diamandis explained on May 16th.

Subscribe now

So, ultimately, I think the questions are how do we prepare students for this new “industrial era” where machines can think and not the last one, and how can individuals and institutions who want to do that be supported?

This is what this Substack is focused on.

Knowledge of the Changing World

A core component of preparing students for the changing world is to understand it ourselves. To support this, in collaboration with Dr. Anand Rao and our AIxHigher Ed podcast, we track weekly updates.

I’ve consolidated emerging trends here.

Coverage of controversies related to data centers, the environment, discrimination, and everything in between is here.

We have a special section related to potential unemployment, universal basic income (UBI), and the restructuring of the social contract here.

Knowledge of AI

Posts throughout this Substack emphasize building basic literacy related to AI and understanding how human and artificial intelligence work (Dasey).

Connecting to Work

We cover what employers are looking for and how to support students as entrepreneurs.

We have a special focus on the “soft,” AKA “durable” skills.

Connecting to Universities

What do universities need to do, and how do they need to restructure? Our posts on that are here.

Whether many universities survive probably depends on whether they focus beyond the four-year degree and reinvent themselves for the community and support lifelong learners.

Practical Action

It’s more than a dream. It’s more than a set of goals.

And it isn’t rocket science.

Classrooms should be transformed beyond a focus on standardized testing and the production of content towards kinetic learning communities that emphasize debate (extensive resources), project-based learning (PBL) (Alcock), interdisciplinary learning, design thinking (Quidwai), entrepreneurship (everyone), robotics, and other forms of AI-Human interaction.

Thank you,

Stefan

stefanbauschard@globalacademic.org
Stefan-bauschard.com

Subscribe now

Subscribe now

Students are graduating unprepared

At the University of Central Florida earlier this month, real estate executive Gloria Caulfield told the graduating class of the College of Arts and Humanities that “the rise of artificial intelligence is the next industrial revolution.”

The booing started before she finished the sentence.

“Okay,” she said, blinking out at the crowd. “I struck a chord.”

A few days later at Middle Tennessee State University, record executive Scott Borchetta — the man who discovered Taylor Swift — told graduates that “AI is rewriting production as we sit here.” When the boos came, he answered them: “Deal with it. Like I said, it’s a tool. You can hear me now or pay me later.”

Then on Friday at the University of Arizona, former Google CEO Eric Schmidt — one of the architects of the world the Class of 2026 is walking into — started talking about AI. The booing began. Schmidt stopped and looked out at the crowd.

“I know what many of you are feeling about that. I can hear you. There is a fear.”

He went on: “There is a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating, that the climate is breaking, that politics are fractured, and that you are inheriting a mess that you did not create.”

He called the fear “rational.” And then he told them the answer was to adapt. To get on the rocket ship. The booing resumed.

This is what the spring of 2026 looks like. The most ceremonial, most carefully scripted moment in American higher education — the one where you stand up and tell young people who they are now — keeps being interrupted by the young people themselves, telling the people on the stage that they don’t want it and that they are unprepared for it.

Ted Dintersmith — the venture capitalist turned education reformer behind Most Likely to Succeed, What School Could Be, and Aftermath: The Life-Changing Math That Schools Won’t Teach You, published this spring — watched these scenes and reached for the only metaphor adequate to them. Whether it takes three years or ten, he wrote, “it’s inevitable that AI will outperform humans in almost all knowledge-worker tasks. Millions of current jobs will vanish… Young adults trained to carry out knowledge-worker assignments will flounder. This is a gigantic issue bearing down on us, and calls for a D-Day mindset.”

Dintersmith is not a tech evangelist and not an AI skeptic. He is the most serious voice in American education reform of the last decade, and he has spent that decade arguing the case he makes in Aftermath: that schools are still teaching kids — in math, and everywhere else — to do exactly what computers now do perfectly. He is telling his own field that it is not responding to the moment.

He is right. And the implication is that this is not a year for tweaks. It is a year to rebuild schools — the curriculum, the assessments, the priorities, the whole architecture — for the world that is actually arriving. The booing is the early warning. What the people on the stage do next is the test.

Subscribe now

What the boo actually says

The students are not confused. They are not anti-technology. Most of them use these tools every day in their coursework — surveys show majorities of college students are using AI in their classes, often with their professors’ encouragement.

What they are reacting to is sharper. They are walking across a stage holding a credential whose value is being repriced in real time, while the person at the microphone — almost always a wealthy older person who built the system doing the repricing — tells them to embrace the future.

A week before Schmidt got booed in Arizona, Jamie Dimon of JPMorgan said the quiet part out loud from Shanghai. The bank, he told Bloomberg, will be hiring “more AI people and fewer bankers.” The 10 percent annual attrition rate at JPMorgan, he noted, affects 25,000 to 30,000 employees a year — enough headroom to shrink the bank-employee category by a third in three years without anybody being laid off. You just stop hiring the new ones. And “I think it will reduce our jobs down the road.”

The same week, Standard Chartered CEO Bill Winters went further at an investor event in Hong Kong, explaining that the bank’s plan to cut 7,800 jobs by 2030 was “not cost-cutting” but “replacing in some cases lower-value human capital with the financial capital and the investment we’re putting in.” When the phrase lower-value human capital hit the headlines, Winters wrote a memo to staff trying to take it back: “Where roles do fall away, it reflects changes in the work, not the value of our people.” The walk-back is more revealing than the original. The plan is the same. The phrase was the mistake.

If Dimon and Winters represent the slow version of this transition — patient, attrition-based, corporate-careful — Zeb Evans, the CEO of the software company ClickUp, represents the fast one. On Thursday, Evans posted a memo announcing he had just reduced his headcount by 22 percent in a single day. He framed it not as a cost cut but as a philosophical reorganization. “The way to operate at the highest level of productivity is changing,” he wrote, “and to win the future, ClickUp needs to change with it.”

The Evans memo is worth reading carefully, because it contains the labor-market thesis stated more bluntly than any CEO has stated it yet. “The common narrative is that AI makes everyone more productive,” he wrote. “It doesn’t.” His claim is that AI makes the best engineers — the ones with the judgment to direct AI agents — radically more productive, and that everyone else slows those engineers down. More code, in his telling, “is just another bottleneck.” He is not saying junior engineers will be less productive. He is saying they will be negative productivity in the new structure. So he cut them. And he announced million-dollar cash compensation bands for the people who remain.

This is the part of the story the gentler corporate framings have been obscuring. The future Evans describes is not “fewer jobs.” It is bifurcated jobs. A small number of people who can orchestrate AI systems get paid spectacularly. The middle of the labor market — the place American education has spent fifty years training young people to occupy — collapses. The bottom of the ladder is removed, as Dimon described. The top of the ladder is raised, as Evans is now demonstrating. The middle, where most graduates have always landed, simply ceases to exist as a category.

Subscribe now

That is the world the Class of 2026 is graduating into. Not a wave of layoffs that makes the news — though Evans just provided one. A wave of non-hires that does not. The bottom rung of the ladder being quietly removed while the speakers on the stage explain that the rocket ship is here and you should be grateful to be near it.

And here is the part that should keep every school leader in America up at night: the students are right that they are unprepared. Their schools did not prepare them. Their schools largely refused to.

Subscribe now

The other side of the lectern

When Demis Hassabis of Google DeepMind closed the I/O keynote this week, he used a phrase no serious AI executive has been willing to say from the main stage until now: we are “standing in the foothills of the singularity.” He called AGI “the most profound and impactful technology ever invented.” And he said, “we’re in a moment of immense promise, but also enormous responsibility.”

Hassabis was talking about the people building the technology. Fair enough. But that sentence applies, with at least as much weight, to a different group: the people who decide what fifty million American K–12 students will learn next year. Superintendents. Heads of school. Principals. Board chairs. State commissioners. Accreditors. Deans. The faculty who train the next generation of teachers.

If Hassabis is even half right about where this is going — and the labor market signals from Dimon and his peers strongly suggest he is — then the people who run schools are sitting on what is plausibly the largest responsibility in the history of their profession. They are deciding what an entire generation knows, can do, and believes about themselves at the exact moment the economic ground underneath knowledge work is shifting.

Subscribe now

Are they treating it that way? Or are they largely running back 2022-3 in 2026-7, with maybe adding a chatbot?

Nearly all schools in 2026 are still preparing students, in the substance of their curriculum and the texture of their assessments, for the economy of 2010. They are still organizing the school day around the assumption that the production of competent written and analytical work product is the central skill the school is paying off on. They are, in short, training young people for exactly the category of work that Jamie Dimon is choosing not to backfill and that Bill Winters is calling lower-value.

This is not a critique of teachers. Teachers are, in my experience, the people in the building most aware of the problem and most starved of permission, training, and time to address it. This is a critique of leadership. It is a critique of the people whose job is to look up from the daily operation of the school and see what is coming.

Subscribe now

The Maginot Line in the curriculum

If Dintersmith is right that this moment calls for a D-Day mindset, it is worth naming the opposite of that mindset, because the opposite is what most American schools are currently practicing.

The opposite of a D-Day mindset is a Maginot Line mindset.

The Maginot Line was the most elaborate, most expensive, most technically sophisticated defensive system in military history when France built it in the 1930s. Reinforced concrete bunkers, retractable artillery, underground rail lines, ventilation systems, mess halls — a marvel of engineering pointed at the German border. It was designed with extraordinary care to fight the war France had just fought, the trench warfare of 1914 to 1918. When the Germans came in 1940, they did not attack the Maginot Line. They went around it through the Ardennes forest, which French planners had judged impassable. France fell in six weeks. The bunkers were still intact. They had been pointed at the wrong war.

This is what most American schools are doing right now. They are perfecting the Maginot Line.

They are still teaching the five-paragraph essay as if essay-writing-as-credentialing has a future, when the credential it once signaled — the ability to produce competent generic written work product to a rubric — is now a commodity that costs fractions of a penny per page. They are still drilling students on procedural math problems that any frontier model now solves in under a second, while Dintersmith’s Aftermath documents the math that actually matters in an AI-saturated world — statistics, probability, estimation, judgment under uncertainty — and which schools largely refuse to teach. They are still defending the AP curriculum, the standardized test, the credit hour, the bell schedule, the four-year degree path as if these structures are load-bearing rather than ornamental.

They are sophisticated. They are defensible. They are pointed at the wrong war.

A Maginot Line mindset is not stupid. It is not even, in its own terms, irrational. The people building it are doing their best with the tools they know, defending what they understand against what they understand. The problem is that the threat is not coming from where the bunkers are pointed. The threat — to the value of the credential, to the relevance of the curriculum, to the entire economic premise of “go to school, learn this content, get a job” — is coming through the Ardennes. It is coming through the labor-market non-hires. It is coming through the collapse of essay-writing as a signal of competence. It is coming through the repricing of every task that schools have been optimized to certify.

The leader who recognizes the Maginot Line for what it is has a choice. The leader who does not has already made one.

What leading would look like

Return for a moment to the Cardona quote at the top of this essay. He is the former U.S. Secretary of Education — the person who held the top federal education job in the country as recently as last year. He is not an outside critic. He is not a tech evangelist. He is the most credible inside voice American education has on this question. And he is telling his own field that the moment we say “AI is the IT department’s job” or “AI is the expert person’s job,” we are failing.

The standard playbook in most schools is precisely the failure Cardona names. Convene an AI committee. Write a use policy. Buy a platform. Move on. That sequence is what schools do when they have decided, without admitting it to themselves, that AI is a technology problem rather than a leadership problem. Cardona is telling them they have decided wrong.

So what does leading actually look like? I do not think school leaders need to become AI experts. I do not think they need to predict which model wins or which company survives. I think they need to do something simpler and harder, which is to ask the question Hassabis answered for himself and Cardona is asking them to answer for their own institutions:

What is the responsibility commensurate with this moment, and am I exercising it?

A leader exercising that responsibility tells families the truth about what their child’s credential might mean in 2030, even when the truth is uncomfortable.

A leader exercising that responsibility audits the curriculum honestly — not for “AI integration,” but for what survives the transition and what could be better inside the transition. Real-time adversarial reasoning survives. Judgment under pressure survives. Ethical deliberation survives. The capacity to read a room, to argue a position you disagree with, to change your mind in public, to negotiate, to lead — these survive. Producing generic written work product to a rubric does not survive. The leader has to be willing to say so.

A leader exercising that responsibility recognizes that the labor market is bifurcating, not just shrinking. The ClickUp memo names what is coming: a small number of people who can orchestrate AI systems and exercise high-trust judgment will be paid extraordinarily; the broad middle in which most graduates have always landed will hollow out. The equity question the school has to answer is not whether to teach AI literacy. It is whether the school is preparing every student — not just the children of professionals — for the upper tier of that bifurcation. The judgment and orchestration capacities that get you into Evans’s million-dollar bands are not the property of any one demographic. They are teachable. They are exactly what debate education has been teaching to students across the income spectrum for a century. The question is whether the school chooses to teach them on purpose.

A leader exercising that responsibility rebuilds assessment around capacities that AI cannot perform for the student. Real conversation. Live argumentation. Defended positions. Performance under questioning. Scaled across the school, with as much emphasis as we put on writing. This is exactly the architecture competitive speech and debate has been running for a century. It is not coincidence. It is the form that emerges when the test cannot be outsourced.

A leader exercising that responsibility creates a role in the school — call it Head of Preparedness, as I have argued elsewhere, or call it something else — whose entire job is to look at what is coming and ensure the institution is adapting at the speed of the change, not at the speed of the committee.

A leader exercising that responsibility stops outsourcing the hard questions to the technology committee and starts owning them at the board table. Because the question of what we are educating children for in a world where Hassabis is correct is not a technology question. It is the central question.

Most school leaders are not doing these things. They are doing some version of the thing the commencement speakers are doing — gesturing at AI, calling it important, urging adaptation, and otherwise running the school the way they ran it in 2022. They are, in effect, asking the next class of graduates to absorb the cost of institutional inertia and call it personal resilience.

That is what the students at UCF and MTSU and Arizona are booing. They just have not aimed the boo at the right target yet.

Subscribe now

A governor moves

On Thursday, while the booing was still echoing through commencement venues around the country, the governor of California issued an executive order.

Gavin Newsom directed his state — the largest economy in the country, the home of nearly every major AI company in the world, the place with the most to gain and the most to lose from this transition — to begin preparing workers and businesses for what he called “the economic disruption” AI is poised to cause. The order enlists state agencies, labor experts, economists, universities, and AI leaders. It calls for worker retraining, employment stability payments, an “AI playbook” for modernizing job programs, a public dashboard showing AI’s impact on the workforce, and policies that “ensure AI advances the public good.”

The line in the press release that should stop every school leader in America cold is this one: “This moment demands that we reimagine the entire system — how we work, how we govern, how we prepare people for the future.”

The governor of California — the politician most exposed to the upside of the AI economy, the one with the most political incentive to be cheerful about it — is using the word reimagine. He is naming all three pillars: work, governance, and preparation. He is, in effect, saying out loud what Hassabis said from the I/O stage, what Dintersmith said with his D-Day metaphor, what Dimon said with his hiring plan, and what the booing students of 2026 said with their voices.

The shock is real. The response has to match the shock.

Newsom can issue executive orders for workforce policy. He cannot issue an executive order for what happens in a fifth-grade classroom in Sacramento, or a tenth-grade English class in Fresno, or a community college math course in Long Beach. That decision is made by school leaders. By superintendents, principals, heads of school, board chairs, deans. By the people the title of this essay is addressed to.

According to Deep View —

If the governor of California can look at this moment and conclude that the entire system needs to be reimagined — work, governance, preparation — what exactly is the school leader in his state, or any other, waiting for?

What we owe the next class

There will be another graduation season next May. There will be more speakers. Some of them will mention AI. The question is whether the Class of 2027 will have any reason to react differently than the Class of 2026 did.

They will only react differently if their schools spent this year doing the work. Not adding an AI literacy module. Not buying a new platform. Doing the actual work — looking at the curriculum, at the assessments, at the daily texture of what students are asked to do, and asking whether any of it makes sense in light of what Hassabis said from the I/O stage, what Dimon and Evans said from their CEO chairs, what Newsom said from Sacramento, what Cardona said from inside the field itself, and what the booing students of 2026 said with their voices.

We can at least try. That is the minimum the moment asks of us. We can try to prepare students for the world that is arriving instead of the one that is leaving. We can try to build curricula around the human capacities that survive. We can try to be honest with families. We can try, as Cardona put it, to treat AI as a leadership lens rather than a technology conversation — and to be the leaders that Hassabis, Newsom, and Cardona were, in their different registers, describing.

If the institutions with public-trust responsibility for educating children will not lead this, the for-profit sector will. It is already moving. The new schools, the AI tutors, the alternative credentialers, the apprenticeship-and-portfolio providers — they are not slowed by curriculum committees or accreditation cycles or the sociology of a faculty meeting. They will not have our values, our equity commitments, or our civic mission. They will have speed, capital, and the willingness to tell parents the truth about the labor market.

The choice in front of school leaders right now is not between change and stability. It is between leading the change with our values intact, or losing the franchise.

Either way, the booing will continue. The only question is who, eventually, deserves it.

Subscribe now

We didn’t have time to record our weekly AIxHigherEd podcast this week, but it’s already been an incredible week.

These are the updates —

Anthropic

Anthropic. Andrej Karpathy — one of the most respected researchers in the field, formerly of OpenAI and Tesla — announced on May 19 that he was joining Anthropic. The line in his post that did the work was this one: “the next few years at the frontier of LLMs will be especially formative.” Karpathy is not a hype account. When he uses the word “formative,” he means it in the sense an embryologist would: the shape that gets set now is the shape that stays.

The same week, the Wall Street Journal reported Anthropic expects $10 .9 billion in revenue for the June quarter — a 130 percent surge — and its first operating profit, projected at $559 million. Yahoo Finance reported a possible trillion-dollar valuation in the next funding round. Anthropic’s revenue run rate at the start of 2025 was barely above $1 billion; at the end of 2025 it was $9 billion; by early April 2026 it was reportedly $30 billion; by May 2026, $44 billion. That is not a growth curve. That is a phase change in how capital flows toward the technology.

There have been so many anti-AI voices on social media posting how AI is not profitable, as if companies aiming for hundreds of billions in revenue would trun profitable in a few months.

Anthropic is also likely to IPO this year. Anthropic retained Wilson Sonsini — the law firm Silicon Valley uses when it goes public — and The Information has reported a target listing date as early as October 2026. On May 12, Bloomberg reported Anthropic was in talks for another $30 billion-plus raise at a pre-money valuation above $900 billion. Secondary markets are not waiting for the official number. Yahoo Finance on May 20 ran the headline that captured the mood: “Anthropic’s shadow IPO market is already flashing trillion-dollar prices.”

OpenAI

Yesterday, OpenAI announced what may be the more consequential item of the week: a general-purpose reasoning model autonomously solved the planar unit distance problem, an open question in mathematics first posed by Paul Erdős in 1946. For eighty years mathematicians believed the best solutions looked roughly like square grids. OpenAI’s model disproved that and discovered a new family of constructions that performs better. The lab framed it carefully: “This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.”

That sentence deserves a re-read. Not “helped a human solve.” Not “assisted in solving.” Autonomously solved. By a general-purpose reasoning model, not a math-specific system. The implication OpenAI drew is the right one: these systems are becoming capable of holding together long, difficult chains of reasoning across distant fields.

Two things happened next that mattered more than the press release.

The first was that Tim Gowers — a Fields Medalist, one of the most decorated living mathematicians, and someone who has thought carefully about AI and mathematics for years — wrote that “if a human had written the paper and submitted it to the Annals of Mathematics and I had been asked for a quick opinion, I would have recommended acceptance without any hesitation. No previous AI-generated proof has come close to that.” The Annals is widely regarded as the most prestigious math journal in the world. A Fields Medalist just said the AI’s proof would have cleared its bar on the first read.

The second was that Po-Shen Loh — Carnegie Mellon math professor, longtime VP of the International Math Olympiad Foundation, and someone who had personally worked on this conjecture — posted publicly that the result will “fundamentally affect the structure of how universities select and retain professors. And more generally the structure of work and creativity.” His framing of the problem is worth quoting at length, because it is the cleanest articulation I have seen of where this leaves higher education:

In college classes, we already have a major issue where take-home assignments are susceptible to students using AI. Even with in-person exams, we’re having issues where students use AI in the bathroom during the exam. Now, how will universities decide which professors to hire and promote? Universities used to judge on the basis of whether you got papers published in highly prestigious journals. Should the person who can use AI to generate massive quantities of publishable results be picked, if it can be done with one-shot prompts?

Read that carefully. A tenured mathematician at one of the world’s top research universities is asking — in public, on LinkedIn — whether the basic criterion that universities have used to hire and promote their faculty for a century still functions. He is not raising this hypothetically. He is raising it because an AI just produced, in one shot, the kind of result that gets a person tenure.

Loh’s own answer is interesting. He says his hiring criteria as an entrepreneur have already shifted. He used to be impressed by competition performance — and this is a man who has coached the U.S. International Math Olympiad team. Now he looks for two things: people who enjoy delighting other people, and people who pursue understanding through their own thought. The skills, he says, follow from those dispositions.

More detail: Remarks on the Disproof of the Unit Distance Conjecture; Noam Brown — Models were only at IMO Gold level a year ago’'; Median for solving most Millennium Prize problems is now 2032.

The same week, OpenAI lined up its IPO. The ChatGPT maker is preparing a confidential filing with Goldman Sachs and Morgan Stanley, with a draft prospectus possibly submitted as early as Friday. Recent private rounds valued OpenAI near $850 billion. An IPO could push that past $1 trillion.

Google Deep Mind

Google DeepMind. At Google I/O on May 19, Sundar Pichai opened by reporting that Gemini is now processing 3.2 quadrillion tokens per month — up from 9.7 trillion just two years ago — and the Gemini app has surpassed 900 million monthly active users, more than double a year ago. AI Overviews in Search hit 2.5 billion MAU and the Gemini app has more than 900 million active users.

The keynote that followed was less a product launch than a survey of how much of daily digital life Google is rebuilding around Gemini in a single quarter:

• Gemini 3.5 across four tiers (Flash-Lite, Flash, Pro (coming soon), Deep Think), with Flash claimed to run four times faster than the previous generation.

On Moonshots Salim identified the strategic split: premium cognition vs. ultra-cheap but fast cognition. Flash serves the latter. And that continuous march of marginal intelligence, cost trending toward zero, has massive implications.

• Gemini Spark — a cloud-based agent that lives on Google’s virtual machines and works in the background, 24/7, syncing across iPhone and Android. The on-stage demo had it managing wedding RSVPs across Docs, Sheets, and Gmail, sending follow-up reminders, and updating a tracker in real time. Google’s framing word for it: “personal agent.”

These agents can do the real work, and Salim saw it: “Today: human → website → cart → checkout. Tomorrow: intent → agent → transaction.” Every CMO will ask: how do I convince a hundred million agents to choose my product? Shopping becomes continuous, invisible, and returnable.

The shift happening right now is much bigger than “AI assistants.” Tech companies are starting to build persistent AI workers that live beside you across your entire digital life.

And slowly, humans are moving from “doing the work” to supervising systems that do the work for them.

• Gemini Omni for video generation and editing — take a photo, get several videos; reframe, re-pose, re-light with conversational prompts. Reviewers on the floor used the phrase “actually looks real.”

• A redesigned, AI-first Search, that now has crossed one billion users. The “ten blue links” era is officially over. Tom’s Guide ran a hands-on piece titled “Google’s new Search is great — unless you’re a journalist,” with the writer admitting “my career is cooked now.”

make me an image for the format in DebateUS theme

The Deep View breaks it down in more detail.

So, yes, it could even, with the right prompt, create an app to teach me Public Forum debate.

• Ask YouTube — Gemini becomes the search index for the largest video corpus on Earth, dropping users at the exact timestamp that answers their question.

• Docs Live, Google Pics, Stitch, Flow Music — voice-first document drafting, AI flyer/infographic creation, “rough idea to working website” generation, and a tool that turns a piano sketch into a professionally mixed track.

• Android XR partnerships with Samsung, Warby Parker, Gentle Monster, and Xreal, putting audio-and-camera-enabled glasses on faces by fall. On-stage demos had the glasses ordering DoorDash coffee, generating Maps directions, and editing photos through voice commands.

Subscribe now

Antigravity 2.0 is replacing Gemini CLI.

And then Demis Hassabis, in a keynote that sat under most of the product news, said the quiet part out loud: “AGI is now on the horizon… When we look back at this time, I think we will realize that we were standing in the foothills of the singularity. It will be a profound moment for humanity.”

On Moonshots, Alex Wissner-Gross described it as: “Every sci-fi trope, everywhere, all at once over the next 10 years… that’s the singularity.”

Capacity

The capacity to serve it all grows. OpenAI's new "Guaranteed Capacity" tier sells “multi-year compute lock-ins” (AWG), with Altman warning the world "will be capacity-constrained for some time." Armada raised $230M to mass-produce modular data centers.

On Tuesday, I appeared as a featured guest on NewsNation to address the accelerating footprint of data centers across the United States and the complex challenges they present to local communities.

The segment, titled “Developing Story: Will America Embrace New Data Tech?”, directly followed the highly publicized story of Coweta County, Georgia homeowner Ansley Brown, who is locked in an intense battle with Georgia Power.

I emphathized with Brown’s plight noted that the data center and power companies have a responsibility to the communities they build in.

“...I do think we need to find ways to build data centers in ways that are the least disruptive to communities. And when the data center companies come in and they build these data centers, they should compensate, they should be taxed at a very high rate... there should be environmental regulations that apply to these data centers. They should have to contribute to the communities...”

Ultimately,, data centers are the infrastructure of the new economy. In the 19th and 20th centuries, America’s economic growth was fueled and defined by the expansion of railroads and highways. Today, data centers and robust power grids are the digital highways driving the future. We cannot have a modern economy without them, but we must be intentional about how they are built.

As utility companies and tech giants push to keep pace with the AI race, finding a balance requires robust public discourse, empathetic policies, and true corporate accountability. Because the country is only at the absolute beginning of the AI revolution, open, empathetic, and serious public debates regarding how this technology impacts people’s homes, lives, and local economies are urgently needed.

Subscribe now

What the week actually means

Three things are happening at once, and each amplifies the others.

Capability is compounding. OpenAI solving an eighty-year-old open problem in mathematics — with a general reasoning model, not a specialist — is the kind of thing that, two years ago, the field’s most cautious researchers said was at least a decade away. The labs are now talking openly about “long, difficult chains of reasoning, connecting ideas across distant fields, and surfacing paths researchers may not have explored.” That is the operational definition of doing real intellectual work. The Gemini Deep Think tier, the Spark agent that “lives in the cloud” and works in the background, OpenAI’s reasoning breakthrough — these are not assistants anymore. They are colleagues with terms of service.

Capital is following the capability and infrastructure faster than it has followed anything in the history of technology. A company barely above $1 billion in revenue at the start of 2025 — Anthropic — is now plausibly worth a trillion dollars. Both burn through training costs that would bankrupt almost any other industry, and investors are lining up to fund the next training run. This is not a bubble in the ordinary sense. It is markets correctly pricing the belief that whoever owns the frontier models in 2028 owns the substrate that every other business runs on.

Deployment is becoming ambient. This is the piece schools are not yet absorbing. Gemini is in the search box for 2.5 billion people. The Gemini app has 900 million users. Spark is a 24/7 agent that lives in the cloud and works while you sleep. The glasses arrive in the fall.

But look closer at the I/O announcements and notice what’s actually happening: Google isn’t building AI products. It is replacing the substrate of the consumer internet with AI, layer by layer. Search results aren’t links anymore — they’re generated mini-apps. Shopping isn’t browsing — it’s an agent in your cart checking part compatibility and executing the purchase. Video search isn’t scrolling through thumbnails — it’s Gemini dropping you at the timestamp that answers your question. Document creation isn’t typing — it’s voice. Image creation isn’t designing — it’s prompting. Website creation isn’t coding — it’s describing. And the interface for all of it is migrating onto your face.

Within twelve to twenty-four months, AI will not be something a student “uses” for an assignment. It will be the medium they think inside — on their phone, in their search, on their face, in the apps they don’t even register as AI products. The interface disappears. The capability stays. The student who isn’t using these tools won’t be the resister. They’ll be the one who can’t afford the $100 Ultra subscription, and that gap will be visible in their work within weeks.

Put those three together and Karpathy’s word is the right one. Formative. The shape we set now is the shape we get.

Subscribe now

Convergence

I want to ask you to do something for the rest of this section: imagine.

Not in the soft, gauzy sense that word usually carries when futurists use it. In the disciplined sense — picture the specific scene, with the specific people, in the specific room, on the specific date. Hold it. Then ask whether anything you know about the trajectory of the technology rules it out.

Imagine an OpenAI reasoning model, running somewhere in a data center, looking at a problem Paul Erdős posed in 1946 — a problem the world’s best mathematicians could not solve in eighty years of trying. It produces, in one shot, a 125-page proof that disproves the conjecture. A Fields Medalist reviews the work and says he would have recommended it for publication in the world’s most prestigious math journal without hesitation.

You do not have to imagine this one. It happened yesterday.

Now imagine a teenager in October, four months from now, walking to class wearing a pair of Warby Parker frames that look like normal glasses. She mumbles a question. A camera on the bridge of the frames captures what she’s looking at. A microphone catches her voice. A cloud agent — Spark, or whatever the equivalent is called by then — receives the multimodal input and reasons across it. An answer comes back through a bone-conduction speaker beside her ear. Three seconds. No app opened. No word typed. She has not even fully finished her thought. The system has already inferred what she meant.

You do not have to imagine that one very hard either. The glasses ship in the fall. Spark was demoed on Tuesday.

Now connect them.

The system answering the teenager’s question and the system that disproved the Erdős conjecture are not different systems. They are different configurations of the same underlying capability — a frontier reasoning model, the kind of thing that lives in the pre-training runs Karpathy was hired to accelerate, exposed through different interfaces with different latency budgets and different harnesses around it. Today, the Erdős-proof model runs internally at OpenAI on long timescales with heavy compute. Tomorrow’s Flash-tier version runs at four times the speed of last year’s, exposed through a search box, a chat app, a pair of glasses. The week after that, a distilled version runs locally on the device itself.

So: imagine being able to casually ask your glasses for the solution to an unsolved math problem. Not a Fermat-class problem on day one — but a research-level problem somewhere on the spectrum between Erdős’s and a homework set. And actually get one.

Hold that scene. A high school junior. Walking between fifth and sixth period. Mumbles, “what about the configurations where the points aren’t on a grid?” — and the glasses give her, in real time, a sketch of an argument that, two years earlier, would have been frontier mathematics.

This is the convergence the labs are racing toward, and it is the convergence schools are not preparing for. The two stories of the week — frontier capability and ambient deployment — are about to be the same story. The thing that just generated a result worthy of publication in the Annals of Mathematics will, within a small number of product cycles, be the thing students casually consult while walking between classes. It will not feel like AGI. It will feel like talking to your sunglasses. The strangeness will be hidden by the ordinariness of the interface.

[Note: The government doesn’t necessarily get it either, asking for 90 days to review new frontier models when new frontier models are being released almost monthly].

Now move the imagining closer to home.

Imagine your own classroom. Not a generic one. Yours. The actual room, the actual desks, the actual student who sits in the third seat from the door and whose name you know. It is two years from now. He is wearing glasses. You have assigned a problem set for in-class work. He looks down at the page. His eyes flick up and to the left for half a second — the small, almost invisible gesture that, in 2027, will mean I just asked. He writes down the answer. He shows no other sign of having received help. There is no laptop open. There is no phone on the desk. There is nothing you can confiscate, nothing you can monitor, nothing you can detect.

Now ask the question. Was that cheating?

The honest answer is that the word cheating depended on a model of school in which the friction between confusion and answer was the point. Remove the friction, and the word stops doing what it used to do. Either we redefine what counts as the student’s work, or we keep using the old definition and pretend we cannot see the contradiction.

Imagine the harder version. Same student. Same room. Same glasses. You ask him a question in discussion — “what do you think about what Aisha just said?” He looks at the wall for half a second. He gives a thoughtful, well-organized answer that builds on Aisha’s point with a connection he could not have made on his own. The class is better for it. The conversation is richer. He is also, in some sense that you cannot quite name, not there. Or rather: he is there, and the machine is there with him, and the boundary between the two has dissolved inside an interface designed specifically to dissolve it.

Is that cheating? Is it collaboration? Is it a learning disability accommodation if every student has it? Is it a new form of literacy, or a new form of dependence, or both at once?

These are not hypothetical questions. They are the curriculum committee meetings of the 2026–2027 school year. They are coming whether the committee is ready or not.

So imagine one more thing, and this is the one I want you to sit with longest. Imagine the version of school that is worth running in a world where the glasses can do all of the above. Not the version that survives by banning the technology — that version will not survive, because the technology will be invisible. Not the version that surrenders to the technology — that version is not worth running. The version in between. The version where the human teacher and the human classroom are doing the one thing the glasses cannot: forming the kind of mind that decides what questions are worth asking in the first place, that knows when an answer is too clean to be trustworthy, that can hold a position under challenge from a real human being who disagrees in real time and will not let it go.

That is the version. That is what school is for, in the configuration that is arriving.

That is the convergence. That is what this week was actually about.

Subscribe now

Now move the imagining closer to home.

Imagine your own classroom. Not a generic one. Yours. The actual room, the actual desks, the actual student who sits in the third seat from the door and whose name you know. It is two years from now. He is wearing glasses. You have assigned a problem set for in-class work. He looks down at the page. His eyes flick up and to the left for half a second — the small, almost invisible gesture that, in 2027, will mean I just asked. He writes down the answer. He shows no other sign of having received help. There is no laptop open. There is no phone on the desk. There is nothing you can confiscate, nothing you can monitor, nothing you can detect.

Now ask the question. Was that cheating?

The honest answer is that the word cheating depended on a model of school in which the friction between confusion and answer was the point. Remove the friction and the word stops doing what it used to do. Either we redefine what counts as the student’s work, or we keep using the old definition and pretend we cannot see the contradiction.

Imagine the harder version. Same student. Same room. Same glasses. You ask him a question in discussion — “what do you think about what Aisha just said?” He looks at the wall for half a second. He gives a thoughtful, well-organized answer that builds on Aisha’s point with a connection he could not have made on his own. The class is better for it. The conversation is richer. He is also, in some sense that you cannot quite name, not there. Or rather: he is there, and the machine is there with him, and the boundary between the two has dissolved inside an interface designed specifically to dissolve it.

Is that cheating? Is it collaboration? Is it a learning disability accommodation if every student has it? Is it a new form of literacy, or a new form of dependence, or both at once?

These are not hypothetical questions. They are the curriculum committee meetings of the 2027–2028 school year. They are coming whether the committee is ready or not.

So imagine one more thing, and this is the one I want you to sit with longest. Imagine the version of school that is worth running in a world where the glasses can do all of the above. Not the version that survives by banning the technology — that version will not survive, because the technology will be invisible. Not the version that surrenders to the technology — that version is not worth running. The version in between. The version where the human teacher and the human classroom are doing the one thing the glasses cannot: forming the kind of mind that decides what questions are worth asking in the first place, that knows when an answer is too clean to be trustworthy, that can hold a position under challenge from a real human being who disagrees in real time and will not let it go.

That is the version. That is what school is for, in the configuration that is arriving.

That is the convergence. That is what this week was actually about.

Subscribe now

What this means for school

I run debate programs. I write about education. I have spent four decades watching schools react to technological change, and the pattern is consistent: many schools wait, then schools panic, then schools bolt the new tool onto the old structure and call it innovation. We do not have time for that pattern this round.

Here is what the week’s news actually demands of educators.

Stop pretending the question is whether to “allow” AI. That question was answered by the 900 million people who opened Gemini this week and the 2.5 billion who use AI Overviews. The students in our classrooms in September will have spent the summer using systems more capable than anything that existed when this school year started. The policy question is not access. It is what we are doing with the time we have students in front of us, given that they will spend the rest of their day inside these tools.

Understand what AI search actually changes. Traditional search required the student to translate a fuzzy question into specific keywords, spell those keywords correctly, choose from a list of ranked links, click through, scan the page, and synthesize what they found across multiple tabs. Every step of that pipeline filtered students by skill. The kid who couldn’t spell “photosynthesis” couldn’t necessarily find the page on photosynthesis. The kid who didn’t know the technical term for what they were looking at had no entry point. The kid who couldn’t read at grade level couldn’t extract the answer from the linked article.

AI search demolishes every one of those filters. Google’s AI Mode now accepts queries in text, voice, image, video, screenshot, or live camera feed — and reasons across all of them simultaneously. A student can point their phone camera at a leaf and ask “what kind of tree is this and why are the edges turning brown?” without knowing the words species, chlorosis, or even tree in English if it’s not their first language. They can upload a photo of a math problem they don’t understand and get a step-by-step explanation in their own preferred reading level. They can speak a question in the rambling, half-formed way kids actually think — “why does it feel like time goes faster when you’re older but also like the days are super long” — and the system uses what Google calls “query fan-out,” generating multiple sub-queries in parallel to triangulate what the student actually meant. Google reports that AI Mode queries are nearly three times longer than traditional ones and that more than one in six US searches are now done by voice or image, with image queries growing forty percent month over month.

Think about what that means in a classroom. For two centuries, the unspoken contract of school has been: we will teach you to decode and encode written language, because written language is the only stable interface between you and the world’s accumulated knowledge. That contract is breaking. The student does not need to spell, decode, or even verbally formulate the question. They can show it. They can mumble it. They can describe the shape of their confusion and the machine will hand back not a list of pages to read but a custom explanation tuned to their level, in their language, with follow-up questions ready. The friction that used to separate the curious from the literate has collapsed.

This is, depending on the day, the most democratizing thing to happen to inquiry since the public library — or the most dangerous, depending on whether students still learn to think when they no longer have to struggle. Probably both. But the relevant point for educators is: every assignment that depends on the student doing the finding — research papers, lookup-based homework, “go investigate and report back” projects — was implicitly grading the student’s ability to operate the old search interface. That skill is now done by the machine. What we measured is gone. What we wanted to measure — whether the student can think with what they find, judge what is true, integrate it into something new — that, we now have to measure directly. Which we have rarely done well, because the proxy was easier.

Stop optimizing for outputs the machines now produce on demand. A well-written five-paragraph essay is no longer evidence of thinking. A correct math proof is no longer evidence of mathematical reasoning. A passable research paper is no longer evidence of research. None of these were ever the point of those assignments — they were proxies for cognitive work we wanted students to do. The proxies have broken. The cognitive work has not. We need new evidence.

Optimize instead for the capacities that don’t transfer. Real-time judgment under adversarial pressure. The ability to listen to another human and respond to what they actually said. The capacity to hold a position, lose it gracefully when proven wrong, and gain something better in the process. Ethical reasoning under uncertainty. The willingness to be confused in public. These are not romantic add-ons. They are the irreducible human contributions in a world where the machine can write, code, summarize, plan, and now, apparently, prove things mathematicians have been stuck on since 1946.

Notice that this is what Po-Shen Loh, watching the same news from inside a top math department, said he is now hiring for: people who delight other people and people who pursue understanding through their own thought. Not competition results — and remember that this is the coach of the U.S. International Math Olympiad team saying competition results have lost some of their signal. Two human-facing dispositions: the social one and the intellectual one. The skills, he argues, follow from those dispositions.

Build the room where those capacities get practiced. I am biased — I coach debate. But the bias is downstream of the analysis, not the other way around. Competitive debate is the only widely available activity in American schools that systematically trains real-time reasoning against a live opponent, in front of a judge, with stakes. It is not the only such activity that could exist. Socratic seminars, structured deliberation, mock trial, model UN, oral defenses of student work, philosophy circles — all of these belong in this category. What they share is that the machine cannot do them for you, because the work is the doing.

Treat the next eighteen months as the window. Hassabis used the word “foothills” deliberately. Foothills are where you decide which mountain you are climbing. The institutional choices schools make this year and next — what we measure, what we reward, what we make room for in the schedule — are the choices that decide whether the next generation walks into the AI economy as participants or as bystanders.

2026-7

Subscribe now

2027-2030

2030+

Subscribe now

This follows up on a recent LinkedIn post by Hassabis —

Subscribe now

This is especially significant because Hassabis’ AGI predictions are “notoriously conservative.”

In fact, only 4 months ago, Hassasbis was predicting it would be longer.

Hassabis’ longer time-frame relative to others always made sense because his definition of AGI is very strict. As he explained a year ago:

What is responsible for the accelerated timelines?

It’s hard to say, but one potential explanation is the Singularity.

The Singularity refers to a hypothetical future moment when artificial intelligence becomes powerful enough to trigger a runaway cycle of self-improvement, leading to changes so rapid and profound that human life is fundamentally transformed in ways we can’t predict from where we stand now. The core idea is that once an AI system becomes capable of designing better versions of itself, each successive generation could be smarter and faster than the last — that improved system then designs an even better successor, and so on, with the cycle compounding so that intelligence improvements which took humans centuries could happen in months, then weeks, then days.

At some point, the curve goes nearly vertical—hence “singularity,” borrowed from physics, where it describes a point where normal rules break down (like the center of a black hole). When Hassabis says "foothills of the singularity," he's suggesting we're approaching the early slope of that curve.

Another potential (and even more likely) explanation or signal is the rapid improvement of the models over the last 6 months. In October 2025, noted researcher (PhD @ Stanford, former Tesla FSD director, OpenAI co-founder, and inventor of the term “vibe coding”) Andrej Karpathy claimed that today’s AI agents were “slop.”

Yesterday, Karpathy announced that he was now heading to Anthropic because he thinks “the next few years at the frontier of LLMs will be especially formative.”

And it’s not just at the development layer. Ken Griffith, CEO & Founder of Citadel, discussed the impact of the improvements of AI models over the last 9 months.

Subscribe now

If Hassabis is right that we’re in the foothills of the singularity, the question for educators isn’t whether to adapt but how fast. Agentic systems now compressing weeks of PhD-level analysis into hours are not a distant horizon — they are the early weather of a climate shift in cognitive work itself.

We have been here before. When America moved from an agricultural to an industrial economy, we didn’t just add factories to a farming society. We built the Common School movement, compulsory attendance, the comprehensive high school, and the land-grant universities — an entire institutional response to a labor market that had fundamentally changed. A child raised to follow the plow needed different capacities than one who would manage a payroll or vote in an industrial democracy.

That ecosystem is now obsolete in its core assumption. For 150 years, schools have been organized around the premise that the scarce and valuable thing is average cognitive output. AGI collapses that economy. When a model can do the median knowledge worker’s daily output at near-zero cost, the premium moves to what AI cannot do and what humans must do regardless: real-time reasoning under uncertainty, ethical judgment without precedent, accountability under cross-examination, the discipline of weighing evidence when both sides are well-argued.

These are not soft skills. They are the hard infrastructure of democratic citizenship and of any economy where humans still decide what the machines should do. They are also, not coincidentally, what debate has been teaching all along.

The agricultural-to-industrial transition took a century to absorb, and it was painful. The industrial-to-AGI transition does not appear willing to wait. We can let it happen to education — in which case schools will be hollowed out by the same forces hollowing out entry-level white-collar work — or we can do what the Common School reformers did: name the new world honestly and build the institutions it requires before the crisis forces our hand.

The foothills are where you choose your path up the mountain. We are there now.

Subscribe now

A student turns in an essay her teacher knows she did not write. The five paragraphs are correct. The thesis sits in the right place. The transitions are smooth. The conclusion restates the introduction without saying so. The rubric is satisfied. The grade is a B+.

Everyone in the room knows what happened. No one knows what to do about it, because the honest answer is the one nobody wants to say.

The AI crisis in education is not that students suddenly learned to outsource their thinking. It is that schools spent a century teaching them to do exactly that — and then panicked when the outsourcing got too efficient.

Plato saw the shape of the problem twenty-four hundred years ago, and he was wrong about it in exactly the way we are now wrong about ChatGPT.

Subscribe now

Thamus was right and wrong

In the Phaedrus, Socrates tells the story of Theuth, the Egyptian god who invents writing and presents it to King Thamus as a gift to humanity. Thamus refuses the gift. Writing, he warns, will not strengthen memory — it will destroy it. People will stop knowing things and start looking them up. They will appear wise without being wise. They will become “the disciples of many things, and will have learned nothing.”

Thamus was right about the mechanism and wrong about the verdict. Writing did hollow out the kind of memory the oral tradition had cultivated. The Homeric bards who could recite tens of thousands of lines are gone, and they are not coming back. But we got something in return — a civilization that could accumulate knowledge across generations rather than losing it with each death. We made the trade. It was the right trade. We just had to give up a particular kind of human capacity to make it.

This is the question Thamus actually pressed, even if he did not know he was pressing it: which cognitive functions can we afford to offload, and which ones, if offloaded, hollow out the human capacity we cannot replace?

Every generation answers this question in practice, by what it builds. Institutionalized education is one of the most ambitious answers ever attempted.

Subscribe now

The first great offload

We tell ourselves that schools exist to develop children. That is true, but it is not what built them.

Mass compulsory schooling — the model nearly every country on Earth now operates — is roughly 150 years old. The 1870 Forster Education Act in England, the Prussian reforms before it, the American common school movement, the wave of national education systems that swept the industrializing world between 1870 and 1920 — none of these emerged because someone discovered childhood. Francisco Ramirez and John Boli’s comparative work on the global diffusion of mass schooling makes this plain: states adopted compulsory education in waves not because pedagogy demanded it but because state-building did. Andy Green’s Education and State Formation tracks the same pattern across England, France, Prussia, and the United States.

What did the institution do, structurally?

It offloaded socialization from families, guilds, and churches onto a centralized state apparatus. It offloaded the transmission of basic literacy from parents who often could not read onto trained specialists. It offloaded time-discipline, deference to written rules, and tolerance for boredom from the household onto a building with a bell.

This was a cognitive and social offload at civilizational scale, and by most measures it was a triumph. Literacy rates that had hovered in the single digits for most of human history climbed past 90% in a century. The institution worked.

Then it kept going.

By the 1910s and 1920s, schooling had taken a second turn — one that matters more for this essay. The institution that received those families’ offload in 1870 was still, in important ways, a human-judgment-driven institution. The teacher decided what to teach today. The headmaster decided who advanced. The community decided what was worth teaching. Then, in roughly fifteen years, that judgment was moved out of human hands and into systems.

Ellwood Cubberley, dean of Stanford’s School of Education and the most influential American educational administrator of his generation, wrote in his 1916 Public School Administration: “Our schools are, in a sense, factories, in which the raw products (children) are to be shaped and fashioned into products… The specifications for manufacturing come from the demands of twentieth-century civilization, and it is the business of the school to build its pupils according to the specifications laid down.” The line is famous because it was offered proudly, not in protest.

Raymond Callahan’s Education and the Cult of Efficiency documented what happened next. Curriculum decisions migrated from teachers to central offices and the new class of professional administrators Cubberley was training. Evaluation migrated from teachers to standardized instruments — the Army Alpha of 1917 proved you could rank-order 1.7 million humans cognitively, on paper, at scale, without ever talking to them, and within a decade the same logic was inside every school district. Sorting migrated from local discernment to tracks set by test scores. Time itself migrated from human rhythm to the bell schedule and the Carnegie Unit. The purpose of school migrated from the local community to a national-industrial frame.

This was the deeper move. The first wave of mass schooling offloaded from families to the institution. The 1920s offloaded from human professionals to systems. Teachers were the first deskilled workforce inside the school, before the students were. The whole apparatus learned to produce cognitive judgments — about what to teach, what was learned, who belonged where, what counted as done — without the cognitive labor of a human exercising professional discretion.

This is the architecture we are still running. It is the architecture AI is now perfecting.

Subscribe now

What we taught the students to do

Schools have been the epicenter of cognitive offloading since the 1920s. Learn what I tell you. Think this. Don’t challenge authority. Cram for the test. Stay within the frame. That is not a caricature; it is the operating system the Cubberley generation installed, and we have been running variants of it for a century. Each upgrade — standards, accountability, NCLB, Common Core, evidence-based curriculum — reinforced the architecture rather than questioned it. The frame tightened. The voice telling students what to think got louder, more confident, more centralized, more “scientific.”

Inside that frame, the student’s job has been clear and consistent: receive, repeat, perform, and do not ask whether the frame itself is the problem.

Consider what the standard apparatus actually trains.

The five-paragraph essay teaches that thinking has a known shape, the shape is given by authority, and conforming to it produces the grade. The student is not asked whether the question deserves five paragraphs, or three, or fifty. The cognitive work of form selection — arguably the central act of any serious writer — has been outsourced to convention.

The rubric tells students in advance what counts as good. This is defended as transparency, and it is transparency, but it is also an instruction to stop asking what good is. The cognitive work of evaluation has been outsourced to a document handed out at the start of the unit.

The textbook presents knowledge as a settled inheritance to be received, not a contested terrain to be navigated. The cognitive work of epistemic vigilance — knowing which claims to trust, which sources are interested, which framings smuggle assumptions — has been outsourced to publishers.

The standardized test rewards the production of pre-correct answers under timed conditions. The cognitive work of defining the problem — the hardest and most valuable thing a human does in any field — has been outsourced to test designers.

And the grade itself is an offload. The teacher does the evaluation; the student receives the verdict. Whether the student agrees with the evaluation, whether she could defend her own work against it, whether she has any internal sense of what is good independent of the mark — none of this is asked. Her job is to receive and adjust.

Paulo Freire named this in 1968. In Pedagogy of the Oppressed he called it the banking model — students treated as containers into which the institution makes deposits, then later checks the balance. Freire’s critique was political; he was watching the banking model train the colonized into docility. But the structure he described is also a structure of offloading. The teacher does the thinking. The student receives the deposit. Half a century later, the system Freire diagnosed has only deepened — and now the depositor can be a chatbot.

We built an institution that ran on offloaded cognition for a century and a half, and then we acted surprised when students offloaded their essays to ChatGPT.

They were doing what we taught them to do. We were just used to being the recipient.

Subscribe now

Schoolishness

I want to press this harder, because the offload of cognition is the visible part — the part you can name in a faculty meeting without losing the room. Underneath it is a deeper offload, and this one is harder to look at directly because the entire apparatus depends on us not looking at it.

The institution does not only offload how students think. It offloads what they are permitted to think about.

It tells them which questions are worth asking and which are not. It draws the perimeter of acceptable inquiry — what is on the test, what is in the curriculum, what is in the unit, what falls under “we’re not going to get into that here.” It tells them which books are on the shelf and, in a rising number of districts, which books have been pulled from the shelf because someone decided the students should not encounter them. It tells them what counts as evidence, what counts as a good source, what counts as a defensible position, what counts as too political, too religious, too sensual, too uncomfortable, too dark, too unsettled.

It tells them, by the architecture of the school day itself, what is worth their attention: forty-five minutes of math, forty-five of English, forty-five of science, a thinner slice of art, almost nothing of philosophy, civics reduced to a test prep unit, ethics absent entirely, the questions that organize a human life relegated to “electives” or to whatever the kid finds outside the building. The schedule is a curriculum. The curriculum is a statement about what is real and what is decoration.

It tells them what their future is for. The college-and-career pipeline is the master frame, and the master frame channels the student’s emerging sense of self toward participation in a specific economic order — productivity, earnings, credentials, employability, fit. Students who do not orient themselves around that frame are diagnosed as disengaged, at-risk, in need of intervention. The institution offloads from the student the cognitive and existential work of deciding what a life is for and replaces it with a pipeline diagram. By the time the student is sixteen the question who am I becoming and why has been quietly substituted with the question what are you applying to. These are not the same question. One is the work of a soul. The other is the work of a guidance counselor.

It tells them which forms of disagreement are tolerated. A student who pushes back on a teacher’s framing in the wrong way at the wrong moment is not having their argument engaged; they are being managed. A student who refuses an assignment on principled grounds is not having a moral seriousness recognized; they are being marked down. The institution preserves the form of disagreement — “controversial issues” units, opinion essays with prompts — but it controls the terms under which disagreement is permitted, and a student who disagrees outside those terms learns quickly that the system does not actually want their mind.

It tells them what to want. The reward structure of the school — grades, honors, advanced tracks, GPA, class rank, recognition, college admission — substitutes an external scaffolding of desire for the internal work of figuring out what one actually cares about. Twelve years inside that scaffolding is enough to permanently confuse the two. Many adults never recover. The high-achieving student who arrives at college unable to answer the question what do you want to study without first asking what looks best on the transcript is not a failure of the system. The system did its job.

This is what Philip Jackson, observing classrooms in the 1960s, called the hidden curriculum — the lessons in deference, punctuality, queueing, gendered participation, and acceptance of evaluation that no syllabus names but every classroom teaches.

It is also what the Black unschooling writer Akilah Richards, in Raising Free People, has been calling, more bluntly, schoolishness — “the things that happen throughout our schooled lives,” the residue of institutionalization that remains in a person’s relationship to their own mind long after the schooling has ended.

Richards’s term is sharper than Jackson’s because it names the thing as a quality of the person rather than a feature of the building. The offload is not only of cognition. It is of orientation. The student is taught not just to look outward for the answer, but to look outward for the question, for the criteria, for the perimeter of permissible thought, for the shape of a worthwhile life, for the verdict on whether their inner life is doing it correctly.

This is also what Ivan Illich saw with unusual clarity in Deschooling Society in 1971 — that the deep effect of schooling is not what it teaches but what it conditions: a population that has learned to confuse being credentialed with being competent, being taught with being learned, and the institutional voice with the voice of reality. And it is what Samuel Bowles and Herbert Gintis, in Schooling in Capitalist America (1976), called the correspondence principle: the structure of school relations mirrors the structure of workplace relations, not by accident but because that is the function of school in a capitalist order. The whole apparatus is calibrated, knowingly or not, to produce the kind of person the economic system can use.

And here is where the AI conversation gets clarifying in a way that almost no one wants to name.

An institution that has spent a century and a half offloading from students the cognitive work of deciding what is worth knowing, what is worth arguing, what books to read, what questions to ask, and what futures to want — is not in a strong position to object when those same students offload the residue of the work onto a language model. The model is the logical endpoint of the apparatus.

We taught students that the answer to what should I think lives outside them. ChatGPT is just a more responsive version of the place we told them to look. We told them their job was to produce the response the institution wanted; AI is exquisitely good at producing the response the institution wants. The students learned the lesson. The lesson was the offload.

This is the part that I think educators have the hardest time saying out loud, because saying it out loud requires acknowledging that the crisis is not a crisis of integrity or of effort or of technology. It is a crisis of what the institution was for. If the institution was for the cultivation of independent minds capable of self-government, the AI moment is a catastrophe. If the institution was for the production of compliant cognitive workers fitted to an industrial economy, the AI moment is just an upgrade in the production line. We have built a system that — for most students, in most classrooms, most of the time — works much better at the second purpose than at the first. We pretended it was doing the first. AI has called the bluff.

Subscribe now

What it did to the teachers

There is a symmetry here that the standard reform conversation refuses to look at.

The same institutional logic that trained students to offload thinking outward has spent the last forty years working, with remarkable consistency, to offload thinking outward from teachers too.

The vehicle is standardized curriculum. The argument is reasonable on its face: teachers vary, outcomes are uneven, and a child in one room should not get a measurably worse education than the child next door because she drew a less effective adult. So we standardize. We adopt a scope and sequence. We script the lessons. We align the assessments. We monitor fidelity of implementation. We narrow the variance.

The unspoken second half of that argument is that we are doing this because we do not trust teachers to make the decisions themselves. If we trusted them, we would not need to script them. The scope and sequence is a vote of no confidence dressed as equity. The institution is saying to the professional standing in front of the children: we will do the thinking. You will deliver it.

There used to be a phrase for this that the reformers said proudly: teacher-proof curriculum. The phrase has fallen out of fashion. The product has not. It has only been renamed — “evidence-based,” “high-quality,” “aligned,” “coherent.” These are good words. They are also, in many implementations, the same offload: the cognitive work of deciding what a particular child in a particular room on a particular morning most needs to think about taken out of the teacher’s hands and handed to a curriculum designer in another state, often working for a publisher, often years before the child was born.

Michael Apple has been making this argument for forty years. He calls it the https://api.pageplace.de/preview/DT0400.9781136634789_A23859914/preview-9781136634789_A23859914.pdf of teaching — the same word industrial sociology uses for what Frederick Taylor and his stopwatch did to craftsmen in the 1910s. The deskilling is structural. It is not a side effect of bad implementation. It is what the design does.

The cost is not borne by teachers alone. It is borne by every student who has encountered a teacher reading from a script with the lights of judgment dimmed, who has watched a thoughtful adult shrug and say the pacing guide says we have to move on, who has asked a real question and received a curricular answer. Children can tell when the adult in front of them has been deskilled. They respond to it by deskilling themselves further.

This is also why AI lands in classrooms the way it does. A teacher whose professional judgment has been displaced for a decade by a centralized curriculum apparatus is not in a strong position to push back when a centralized AI apparatus arrives to displace what remains. The institutional muscle for no, I will decide what this child needs has atrophied. Teachers were already being told to offload their thinking to the program. Now the program is offloading itself to the model. The students are doing what the teachers were doing. The teachers are doing what the institution had always done.

There is no version of this where children learn to think under pressure in front of a teacher who has been instructed not to.

Subscribe now

The credential collapse

Step back further.

The whole stacked system of K–12 grades, transcripts, GPAs, standardized tests, degrees, and credentials is itself a vast cognitive offload — this one from society onto institutions.

When an employer hires a bachelor’s degree, she is outsourcing the judgment of competence to the institution that conferred the degree. When a graduate program admits a 3.9 GPA, it is outsourcing the judgment of intellectual quality to the previous institution’s grading. When a state licenses a teacher, a nurse, a lawyer, a doctor, it is outsourcing the judgment of trustworthy practice to a credentialing apparatus. The credential is a shortcut. It exists because individually evaluating every person for every position would be impossibly expensive. So we built a machine that evaluates once, stamps once, and lets the stamp travel.

This is the conflation Illich named half a century ago — schooling gradually replacing learning as the thing society could verify, until eventually the verification was the learning, at least for the purposes of the labor market. The credential became the substance. The stamp became the thing.

What AI is doing to that stamp is roughly what the printing press did to the scribe’s monopoly. The credential rested on a signal — that the holder could do something the unstamped could not — and AI is collapsing the signal across a widening range of cognitive tasks. If a paralegal’s core work can be done by an AI system at one-twentieth the cost and three times the speed, the credential does not vanish overnight. But its premium collapses. The wage was paying for a scarcity that no longer exists.

The institutional response so far has been institutional self-defense — moratoriums, bans, AI-proof assessments, return-to-blue-book proctoring, accreditation regimes that treat AI as a threat to academic integrity rather than as a hammer falling on the credential’s foundations. None of it works. None of it can. You cannot defend a credential whose signal has collapsed by tightening the rules of a test that no longer measures the thing.

The serious question is not how to preserve the credential. It is what we can certify after the credential. What becomes worth measuring when the old measurements no longer mark a difference.

Subscribe now

Where the student cannot disappear

The question is not whether students will use AI. They will. The question is where the institution creates moments in which the student cannot disappear behind the tool.

This is the question the next decade of education policy will be organized around, whether the field knows it yet or not:

Which human cognitive capacities, if offloaded onto AI, end the thing that makes a citizen a citizen?

When you press on this, the answer is not abstract. It picks out a recognizable cluster of capacities — the same cluster, as it happens, that has been cultivated for two and a half millennia by one of the oldest pedagogical traditions in the West.

Real-time adversarial reasoning. Another person is sitting across from you, attacking your argument, looking for the weak link, pressing you on the assumption you tried to hide. There is no offload. The model cannot stand in for you in the cross-examination. You are responsible for the next sentence, and the next, and the one after that, and you have to produce them in a body that is being watched.

Ethical judgment under uncertainty. The data is incomplete, the stakes are real, the principles conflict, and you have to act anyway. No system offloads this for you, because the act of judging is what makes you the agent rather than the executor.

Accountability. Standing behind a claim. Owning it when it fails. Revising in public. This is not a skill the way “five-paragraph essay” is a skill. It is a posture toward truth, practiced in front of others, or it does not develop.

Defending ideas under cross-examination. This is the formative test the system has not figured out how to fake, because the test is the live performance of a mind, observable by other minds, in conditions where the production of the next utterance is structurally non-offloadable.

These are, not coincidentally, the capacities that form the spine of debate as a pedagogical practice — Dialogue, Evidence, Balance/Argumentation, Thinking Under Pressure, Ethics. The D.E.B.A.T.E. framework is not a repackaging of debate as something fashionable. It is an attempt to name precisely the cluster of capacities that cannot be put on the conveyor belt without ending the citizen.

It is also the most AI-resilient formative assessment we have. A live cross-examination is not gameable by ChatGPT. A debate round is not generated by Claude. An adversarial exchange in which the student has to think on her feet in front of a judge who will press her is a form of measurement that has survived every previous wave of cognitive offloading — the printing press, the calculator, the internet, the search engine — and will survive this one.

The reason is structural. In a debate round, there is no rubric standing between the student and the act. No five-paragraph form to hide behind. No answer key. No textbook. There is another human, watching, listening, pressing. The student is responsible for the next utterance. She cannot offload it. If she has not done the thinking, the room knows.

That is what we cannot afford to offload. That is what is left when the rest is gone.

Who steps up

The serious work begins after the diagnosis.

If institutionalized schooling was the great cognitive offload of the industrial age, and if AI has now exposed what that offload cost us, then the question is not how to preserve the system. It is what comes next — and who builds it.

Start with what is obvious and what the critique cannot pretend away. Students still need content. They need to know things. They need to have read books, encountered history, learned mathematics, understood the methods of science, struggled with ideas they did not choose, met traditions older than themselves. The argument has never been against content. The argument is against the delivery mechanism the industrial school built around content — the standardized curriculum, the pacing guide, the scripted lesson, the rubric, the receipt-and-repeat assessment cycle that trained students to relate to ideas as inventory rather than as terrain.

But there is a deeper version of this argument, and it is the one most reformers will not say out loud. It is not just the delivery mechanism. It is the obsession the mechanism rests on. It is the conviction that there is some right body of content most students need, that they should spend 8 hours a day on it, that we can specify it in advance, that getting the specification right is a solvable problem, and that the next reform cycle — better standards, better materials, better alignment, better implementation — will finally deliver it. This conviction is the engine of standardization. It is what keeps the apparatus running through wave after wave of evidence that it is not working. And it is held with a quasi-religious intensity by people who would not call themselves religious about anything else.

The obsession has three layers. The first is the belief that most students need the same content, even as they age — that what an seventeen-year-old in Boise needs to learn is also what an eleven-year-old in Atlanta needs to learn, which is also what an eleven-year-old in Honolulu needs to learn, and that the differences between them are noise to be filtered out rather than information to be honored.

The second is the belief that there is some magic content — that if we could just identify the right canonical knowledge base, the right cultural literacy list, the right sequence of skills, the right grade-level expectations, the students would learn it and the system would work. Every generation of reformers believes they are close to this list. Each generation produces its own version. None of the versions has produced the outcomes claimed for it. And the third is the belief that the adults who design the content know more than they actually do — about which knowledge will matter, in which order, for which children, for which futures, in a world none of us can yet see clearly.

None of these beliefs survives serious contact with what AI is now doing to the labor market, the information environment, or the production of cultural knowledge. The futures the standardized curriculum was designed to prepare students for are dissolving in real time. The content the curriculum was designed to deliver is now available to any student at any moment, for free, in any language, on demand. The adults specifying the curriculum cannot reliably tell you which capacities a person born in 2018 will most need in 2040, because nobody can. And yet the standardization conversation continues — at the federal level, the state level, the district level, the school level — as if specifying the content harder will finally make the system work.

It will not. The obsession with getting the content right is the institutional version of the offload the essay began describing. It is the apparatus trying to do the thinking that only the student, in conversation with adults who actually know her, can do. The standardized curriculum is not the answer to a difficult question. It is the institution refusing to ask the question.

Content does not have to be standardized (or only standardized) to be serious. In fact, the more standardized it becomes, the easier it is for AI to replace the human who delivers it. This is the structural irony of the moment we are in. A society that produces standardized minds — minds shaped by the same scope and sequence, evaluated by the same rubric, certified by the same credential — is producing exactly the population AI is best at substituting for. The factory model is not a defense against AI. The factory model is what AI is replacing. The more our students look like the average graduate of the average program, the more replaceable they are.

The way out is not less content but less standardization of content. Different schools, different traditions, different paths through the material, different combinations of subjects, different ways of measuring what a student has actually come to know and do. Variance is a feature, not a bug, in an age when the median performance of any cognitive task is about to be free.

And this is where the argument has to turn from the institution to the student.

There is a story the industrial school tells about students, and it is one of the most expensive lies in education. The story is that students are not naturally interested in learning, that interest has to be manufactured, and that the school’s job is to motivate the unmotivated through grades, gold stars, threats, rewards, behavioral incentives, and the long-term promise of credentials. The story is necessary for the system to function. If students were genuinely interested in what they were being asked to do, the entire apparatus of compliance — the bell, the hall pass, the late penalty, the missing assignment notification, the GPA — would be redundant.

The story is also false.

Students have interests. They have many of them, often passionately, often in defiance of every effort to redirect them. They have interests in music, in games, in social dynamics, in fashion, in sports, in politics they’re told they’re too young to care about, in identities they’re told they’re too young to claim, in injustices they see clearly and adults pretend not to. They make things, learn languages, master complex technical systems, debate online with strangers, build communities, teach themselves skills the curriculum does not offer. They do all of this while also being told they are disengaged from school. The disengagement is real. It is not a disengagement from learning. It is a disengagement from the industrial form of learning the school is offering them.

What the industrial school stripped out, more than any specific content, was agency — the experience of one’s own choices having consequences. Agency is not the same as motivation. Motivation is something teachers try to install in students. Agency is something the student already has, that the institution either makes room for or grinds down. The standardized classroom is a low-agency environment by design. The student does not choose the topic, the text, the pace, the format, the question, the audience, the timeline, the partners, or the standard of judgment. She is told all of it. After twelve years of being told, she develops what the industrial school then diagnoses as disengagement — and what is actually the rational adaptation of a person whose agency has been systematically denied. She has learned that her preferences do not move the system. So she has stopped offering them.

A school that develops agency is not the industrial school with better engagement strategies bolted on top. It is a structurally different relationship between the student and the work. The student helps decide what is studied, when, with whom, to what end, and by what standard. The teacher is no longer the dispenser of approved content but the more experienced practitioner working alongside her, raising the level of what she is trying to do, defending the discipline’s traditions where they matter, getting out of the way where they do not. The standard is not the rubric. The standard is whether the work is any good — judged by adults who have actually done the kind of work being judged, by other students whose own work has earned them the right to a vote, and, eventually, by the student herself, who has developed taste through enough exposure to know.

This is harder than the industrial model, not easier. It requires adults willing to give up control they could legally exercise. It requires institutions willing to admit that their authority over what a student studies is no longer self-evident. It requires the recognition that interest, once permitted to be real, will sometimes lead the student to places the curriculum did not anticipate — and that this is the point, not a problem. Most institutional resistance to student agency, dressed up in the language of standards and rigor, is actually resistance to the loss of adult control. It will have to be named and faced.

The students are not the obstacle. The students are the engine. The obstacle is a century of institutional muscle memory that does not know how to operate when the student is allowed to want something.

This points at what the redesign actually requires. Not all new tools. Not better technology in current classrooms. The redesign requires new practices — different ways of learning, organized around capacities the machine cannot replace, with the student’s own interest as the live current running through them.

None of this means there is nothing all students should learn. It is almost certainly true that some shared content matters, particularly in the lower grades — that a society which fails to teach every child to read, to do basic arithmetic, to recognize the rudiments of how the world works and how its own past was shaped, has failed them in a way that cannot be repaired later. The early years carry an obligation the rest of schooling does not.

But even there — even where a shared floor is necessary — the institution has to make room. Room for the student to pursue what she finds, room for the questions the curriculum did not anticipate, room for the work she chooses to do well beyond what was assigned, room for the interests the school has historically diagnosed as distractions.

Agency is not the reward for finishing the standardized portion. It is the soil the standardized portion is supposed to be planted in. A school that teaches a child to read but does not let her read what she wants has taught her the technique and revoked the point. A school that teaches the methods of science but does not let her ask questions the curriculum cannot grade has trained a technician and refused her the inheritance. The shared floor is real. So is the human standing on it.

This points at what the redesign actually requires. Not new tools. Not better technology in classrooms. The redesign requires new practices — different ways of learning, organized around capacities the machine cannot replace, with the student’s own interest as the live current running through them.

Some of them are old. Apprenticeship is old. So is the seminar, where students sit around a table and argue about a text. So is the workshop, where someone who knows how to do something teaches someone who does not, in person, over time, with judgment passing both ways. So is debate. So is the studio, the lab, the rehearsal room, the field trip with a guide who knows the territory. These forms cultivated humans for centuries before the factory model arrived. They worked then. They will work now. They were displaced by the industrial school not because they were ineffective but because they did not scale to mass literacy in the way the bell schedule did. The conditions that made the trade-off worth it are ending.

Some of them will be new, and we do not yet know what they look like. The next form of education is going to be invented in places the current system does not look. It will be invented by teachers who refuse the script, by parents pulling kids out of systems that no longer serve them, by debate coaches and writing instructors and music teachers and shop teachers who already know what non-offloadable work feels like, by employers who get tired of waiting for credentials that mean nothing, by community institutions remembering what they used to do, by students inventing things their teachers cannot grade. It will not look like a single national program. It will look like a hundred experiments, most of which fail, some of which root, and from which — if we are honest about what we are seeing — a different kind of institution slowly comes into focus.

The question is who steps up. Not who is appointed. Not who is funded by the next federal initiative. Who, in the actual places where students and adults meet, takes responsibility for the work that the industrial school can no longer pretend to do.

And the redesign cannot be optimized for the AI economy. That is the trap. If we treat school as a workforce-preparation problem, we will lose the argument, because AI is already a better workforce than most of our graduates. The redesign has to be optimized for the kind of human a person becomes inside it: someone capable of judgment, capable of taste, capable of ethical reasoning, capable of standing behind a claim, capable of caring about something other than her own credentialed advantage, capable of participating in a democracy that will not survive a population that has outsourced its thinking.

That is human flourishing in the AI age. Not preparation for the economy. Preparation for the life the economy is supposed to serve.

This is the work. It is not glamorous, it is not centralized, and it will not be funded by the people who funded the last hundred years of schooling. It will be done by people who notice that the children in front of them deserve better than what the system is currently producing, and who decide to build something else, often without permission.

The institutions that figure this out first — schools, networks, leagues, programs, communities — will not be the ones that defended the credential the longest. They will be the ones that asked, earliest and most honestly, what a young human is for.

Subscribe now

Coda

Thamus was right about the mechanism. Every offload hollows out the capacity it replaces. The bard’s memory is gone. The mental arithmetic that filled Victorian schoolbooks is going. Cursive is going. Whatever cognitive practice the textbook trained — the patient receipt of an authoritative narrative — is going too.

The question is never whether to offload. We have been doing it since Theuth presented his gift, and we will not stop.

The question is what we hold back from the trade because losing it would cost us something we cannot recover.

Institutionalized education was the great offload of the industrial age. It worked. It also taught two generations to offload their own thinking onto the structures designed to evaluate them, deskilled the teachers who were supposed to know better, and built a credentialing apparatus that offloaded society’s judgment onto a stamp that AI is now removing the ink from.

We do not need to defend the old structure. We need to be honest about what it was, what it can no longer pretend to be, and what we will not let go of next.

Cross-examination. Live argument. Defense under pressure. Judgment in front of other minds.

A student, in a room with other students, defending an idea she has actually thought.

That is the part we keep.

Further reading

The arguments here draw on a long lineage. For readers who want to follow any of the threads:

• Plato, Phaedrus, c. 370 BCE — the original argument that every new technology of mind is also a loss.

• Francisco O. Ramirez & John Boli, “The Political Construction of Mass Schooling,” Sociology of Education, 1987 — compulsory schooling as a project of state-building, not pedagogy.

• Andy Green, Education and State Formation, 1990 — the comparative case across England, France, Prussia, and the USA.

• Raymond E. Callahan, Education and the Cult of Efficiency, 1962 — the canonical history of how Frederick Taylor’s scientific management was imported into American school administration in the 1910s and 1920s.

• Philip W. Jackson, Life in Classrooms, 1968 — the original hidden curriculum argument.

• Paulo Freire, Pedagogy of the Oppressed, 1968/1970 — the banking model of education as political domestication.

• Ivan Illich, Deschooling Society, 1971 — the credential / learning conflation.

• Samuel Bowles & Herbert Gintis, Schooling in Capitalist America, 1976 — the correspondence principle: school structure mirrors workplace structure because that is the function. Revisited.

• Michael W. Apple, Teachers and Texts, 1986 — the deskilling and proletarianization of teaching under standardized curriculum.

• Akilah S. Richards, Raising Free People, 2020 — the contemporary articulation of schoolishness as a quality of the person.

TL;DR

Dr. Theresa Burriss, keynote speaker for UMW’s upcoming Reimagining the Liberal Arts in the Age of AI conference, frames the AI–humanity relationship through the metaphor of “mending” rather than resolving — using story rather than argument as the entry point to hard conversations. Drawing on Appalachian studies, feminist theory, and existentialist philosophy (with a Derridean preference for “both/and” over “either/or”), she argues that productive tension is unavoidable and necessary: between authenticity and artificiality, unity and diversity, nature and nurture, AI as tool versus AI as thought partner.

On pedagogy, she can tell when students offload entirely to AI (”vapid rhetoric, no voice”) but acknowledges sophisticated prompting can yield genuinely authentic output, complicating older definitions of student work.

On representation, she warns that AI replicates and amplifies centuries of stereotyping of marginalized communities like Appalachia, citing a case where faculty generated fake Appalachians via ChatGPT instead of engaging real ones.

On environmental justice — informed by her seat on the Virginia Council on Environmental Justice — she connects data centers to Appalachia’s long history as a “national sacrifice zone,” noting that whether they help or harm a community depends almost entirely on local leadership and civic participation.

She defends the liberal arts as training in critical and creative thinking and in decolonizing the imagination (drawing on bell hooks), and sees a democratizing role for AI in giving voice to people with expertise but not writing skill. She rejects both AI absolutism and AI refusal, advocating instead for engaged middle-ground negotiation. Her hope: that people are pushing back on for-profit homogenization, and that E.O. Wilson’s biophilia — the innate human pull toward the real and the natural — will persist alongside the artificial.

Subscribe now

Welcome and Conference Context

Anand: Well, welcome back everybody. We’re really excited to talk today to Dr. Theresa Burriss, who’s an independent multidisciplinary Appalachian studies scholar and educator. Thank you so much for joining us today, Theresa.

Dr. Theresa Burriss: Thank you very much for having me. I’m looking forward to speaking with you all.

Anand: Great. Thanks. And I’m really excited to have this discussion because I’m looking forward to welcoming you to UMW later this summer in July for the Reimagining the Liberal Arts in the Age of AI conference, and you’re one of our keynote speakers, which I’m looking forward to.

Dr. Theresa Burriss: Thank you. I was very honored when Brent Blevins invited me to speak. So, hopefully, I will contribute to the discussion.

Anand: Well, I remember her telling the rest of the organizing committee about seeing you speak at another conference and how happy she was with that discussion and really wanted to bring that to campus. So, I’m glad we have a chance to do that.

Subscribe now

“Mend a Story”: The Origin of the Keynote Title

Anand: I want to ask you a little bit about that keynote so that we can kind of whet the appetite of everybody that’s watching so they have a sense of what to look forward to, and tell us a little bit about the title because the title starts with “Mend a Story.” What I love about that is that both of those words, mend and story, are really pulling in some really interesting perspectives already—to think about mending rather than even reconciling and trying to recapture or correct something, and then thinking of it as a story as opposed to an argument. Tell us a little bit about that. How do you put that together and how do you envision it?

Dr. Theresa Burriss: Well, I love that you have picked up on that. The symbolism is important, but let me say this: Mend is actually the title of a painting that my artist friend, Suzanne Stryker, gifted me, and it is of a spider in the middle of a web. The title of her painting is Mend.

But it is very fitting in this context, like you said, of there seeming to be some kind of friction or disruption, if you will, and we are mending it. But the way that I started thinking about this and reconciling authenticity and AI, I used Suzanne’s painting as a launching pad. Instead of writing her a traditional thank-you note, I decided to write her a poem on my own without using AI.

Well, my psychologist husband—who is not a creative, he is a scientist—used AI and fed it with prompts. He used Copilot and generated a poem to give to Suzanne. So, this is where this actually originated. But I think that to your point, I mean, that’s really mending the story. As an Appalachian studies scholar, I feel like stories are a way to invite people into conversation in a non-threatening way, as opposed to an argument. And so it really fits in both capacities to demonstrate where AI fits in our lives.

Anand: I love that sense of maybe even fitting AI and humanity into this web. That’s something to explore. I like that.

Dr. Theresa Burriss: Definitely. Yeah.

Subscribe now

Living in Tension: Authenticity, Artificiality, and Existentialism

Stefan: And you mentioned this a bit with authenticity. But I know in the keynote, it sets up kind of the authenticity and artificiality tension there. But it’s not something to overcome. It’s just like, well, you know, we can kind of live in this tension. And maybe there’s a question of how to live well within the tension. Do you have any kind of thoughts or reflections on that?

Dr. Theresa Burriss: I very much do. In fact, one of the things that I talk about with Suzanne’s painting—yes, it is an authentic painting, but it is not the web and the spider in and of themselves. It is a representation of that. So, even that is problematized: the authenticity of that web and spider.

And Stefan, to your point, I think that all of us live with tensions in our lives, whether we name them as such or even acknowledge them. One of the issues that I talk about in particular—and it derives from feminist theory, but I use it in an Appalachian context—is this tension: What does it mean when we live with—I live with the tension of Appalachia in the singular and Appalachias in the plural. Because going back to feminist theory, there is political might, if you will, in unity and in rallying around a particular identity, but in that process, the diversity within that unity is lost. So, I think that here again, whether humans overtly acknowledge it or not, we live in constant tensions. I think that some people want to avoid that because it’s uncomfortable, but it is inevitable, is my thought.

Anand: That discomfort is a really interesting aspect of it. Living in tension, or existing in that kind of tension, sometimes is really productive, as uncomfortable as it might be. And that’s another aspect of this when we’re thinking about authenticity that I wanted to ask you about. You reference in the abstract about the keynote talking about existentialism and thinking about even kind of juxtaposing or presenting different positions there from Søren Kierkegaard to Nietzsche. There’s some natural tension there between their perspectives as, you know, maybe living more with a religious self before God versus a self-creation from another perspective. Do you think that maybe there’s one of those accounts of authenticity that sits better with you, or that you think might be more useful? Is there a benefit of having that tension between the two as we start to explore it?

Dr. Theresa Burriss: So, I default to Derrida and deconstruction: that it is typically a “both/and,” rarely is it an “either/or.” Because I was a philosophy major at Emory University and took an entire semester seminar on Nietzsche, I lean a bit more towards Nietzsche just in certain arguments, I will say.

But I think that all of the existentialists invite us to consider our place in the world, and each person has to determine what philosophy resonates with them the most. I mean, Kierkegaard is certainly going to resonate stronger with some of my friends who are very much religiously aligned, whereas with Nietzsche—I mean, I will say, thinking about Nietzsche and “the herd,” here again, I studied him ad nauseam.

I used to be—I was just talking to somebody recently about this—I used to be a strict, strict existentialist in the sense that I thought you were born as a blank slate. Then I became a mother, and I realized that was folly. There is definitely—talk about tension—there is a tension between nature and nurture. My two sons definitely came into this world prescribed certain ways, so it didn’t matter what the nurture was, the nature was there. So, even a philosopher is going to complicate it. And, you know, I mean, what does Walt Whitman say? “Do I contradict myself? Very well then, I contradict myself. I am large, I contain multitudes.”

Subscribe now

Story vs. Argument: The Pedagogical Entry Point

Anand: Let me ask you a little bit about this because I think you’re right. There are limits to that tabula rasa approach that we could just be a blank slate. Stefan and I both studied a lot of philosophy. I was an undergraduate philosophy major as well, and studied rhetoric. What I always enjoyed about that was the sense of that inherent contradiction, the conflict, the discomfort.

So, I want to go back to something you said earlier about why it’s story versus argument, because there’s an access point, and there is an attempt at an access of comfort to be able to welcome people in. Not many philosophical discussions are that welcoming, necessarily, or that comfortable. So, I’m wondering how you view that tension even between the entry point and then when you get to the heart of it and how you start to work through it. This is something that we often experience where our backgrounds are in debate. We were actually debate partners in college, and there is that worry from some people that debate is just confrontational and argumentative, and we want something that’s more consensus-based. But even when you offer that entry point of something consensus-based or not as confrontational, there’s still some conflict there with trying to resolve something.

Dr. Theresa Burriss: And I think that that is where the growth occurs for multiple parties. Where my mind went was, as a lifelong educator, it is my goal to create a safe space for my students. I don’t want any student to feel alienated. Now, I do hold accountability that there is civil discourse; you treat others with respect, especially when you have differing opinions. But it’s incumbent upon the educator to create that safe space in the classroom. That’s the opening; that’s the story opening. And then once that is there, that is where we can begin to engage.

Debate does not have to be negative. I mean, I’m a law school dropout. I went to law school for one year, and it wasn’t because of the debate that I didn’t like it; it just wasn’t my career path. In fact, I enjoyed that and I still enjoy that. I like to be challenged, and I think if I mirror that to my students, it works. I’ve witnessed that over 29 years, if that makes sense.

Subscribe now

Student Voice and the Authenticity Continuum

Stefan: Great. Thank you. You know, you talked about teaching a couple of times, and we talked about authenticity. When students come into the classroom now, a lot of them have been using or are familiar with some of these generative models for a while. Not only is their output impacted by that, but kind of their own conversations—whether they’re textual or otherwise—with these AIs.

It’s kind of hard to say because it’s maybe more in the tension, but do you feel their voice is different, maybe in both a literal sense but also just in a more figurative sense? Like, how they’re expressing themselves or what they’re thinking? Do you see how the language models are impacting whatever we want to define as the authenticity of the voices, even if it’s just your intuitive sense?

Dr. Theresa Burriss: I can definitely tell when a student—I look at the use of AI on a continuum, with no AI usage on one end, a lot of things in between, and total AI reliance on the other. I definitely can tell when a student has completely and holistically relied on AI because often it’s vapid rhetoric. There is no voice. I think because I’ve been an educator for so long and a lifelong reader, you know when the voice is missing.

Now, I will say, as students get more sophisticated—this is something that maybe we’ll talk about in a bit—the input, what humans input into the AI engine, determines so much of what comes out. I have read some things where there’s been more of an intentional, creative input, and then the outcome seems more authentic. But here again, I mean, that too is problematic.

To get to your point of how I help students think about their voice, I don’t ever want to give up the importance of discussion. In the classroom, face-to-face—even when I was teaching in a doctoral program in education, it was through the computer because it was virtual—those conversations are critical, even for us as adults, because our voices are evolving as we’re changing as humans, hopefully. So, our voice changes. But I think that here again, providing that safe space to explore who they are and engage in conversations with each other as peers, and also with me, is critical.

Pedagogical Vertigo: When the Tools Get Good

Anand: That idea of what authenticity means anymore, I think, is part of what is throwing a lot of faculty off. I talk to a lot of colleagues who are experiencing a kind of pedagogical vertigo, where the ground’s not really stable underneath us anymore. Some people might say that they can tell that something’s written by AI, but then they see something where a student has it totally written by AI, but the student put a lot of work into the prompting, discussion, and preparation of materials. Then you think, okay, well, then they did do a lot of work, and it doesn’t seem vapid, and there’s some real qualitative value to this output, but it’s not what we would have considered authentic student writing before. Does that mean it’s inauthentic? No, it’s a different kind of authenticity. And so that’s something to negotiate and kind of manage.

I’m curious how you see that perhaps going, especially as we’re looking at both what you’ve already identified as students developing a better facility with the tools and learning how to use them maybe more responsibly because they’re thinking, “I’m really putting a lot into this,” but then it’s writing it, as well as the tools just getting so much better. I mean, we were joking around a little bit about the use that I was making of Claude, and I often talk about how I’m using Claude in my writing, preparing questions for interviews, and developing things. I feel like it’s expanding my capabilities in some important ways. But it certainly makes me question: well, if it’s different than what I would have done on my own, does that mean it’s inauthentic, or does that mean it’s a thought partner? Can I view it as just a tool anymore if it’s contributing in that way? It seems really complicated. How do you negotiate some of that?

Dr. Theresa Burriss: I think it is a constant negotiation. Even when you get to that output and it does appear to be more authentic, however we’re defining that—and that is such a difficult term to define and agree upon—I will say there’s still—I encourage human intervention even after the output. Yes, I do encourage that.

To your point, AI is a tool; that is my belief. It does not replace humans; it cannot replace humans. I know there are a lot of doomsday sayers that will say we’re going to be run by robots. And as a runner—I don’t know if you all just saw that a robot just got the half-marathon world record.

Anand: We just talked about that a couple of weeks ago in one of our updates, actually. Yeah, that was something.

Dr. Theresa Burriss: Okay. Well, I will never run that fast, so, and I will never generate text as quickly as AI. As humans, we have limitations, and technology—whether it’s in the medical field, or education, or the space race, whatever—technology is to enhance what humans are capable of doing.

And so, I see AI as an important disruptor that very much needs to be monitored with guardrails, with critical thinking, ethics, and morals. Because we know, and we’ve already witnessed, how it can be used for extreme harm and ill. To the point of a young man in California committing suicide—I listened to a gentleman who talked to our Rotary Club in Bristol about the dangers of young people and chatbots, and who’s controlling them. There are—I’m not telling you anything new—very sinister people out there.

One of the things that I did want to mention talking about young people—and I can’t remember the study right now—but it was about how AI is dangerous to young, developing minds because those synapses that fire in what we know as learning don’t fire when AI is doing all the work. So here again, I think these conversations are complicated. We’re still figuring things out. What role does it play at different age levels?

Anand: Yeah. And I guess thinking about that role of where it goes, even if it is developing much further, maybe it’s that difference between replacement versus displacement. And so it’s not thinking that we’re going to not have a place, but it’s going to be a very different placement for humans, depending upon what we do with it. That’s one of the things that I’ve heard employers say: AI is not going to replace human jobs; it’s humans who don’t have AI literacy that will be replaced.

Subscribe now

Representation, Appalachia, and Algorithmic Stereotyping

Stefan: Yeah, it seems like a necessary condition, even if it may not be a sufficient condition, in the workforce these days, especially moving forward or for new hires.

But the other thing I want to ask about—I know a lot of your scholarly work is grounded in Appalachia and your interests there. Of course, these are related: there are questions of how Appalachia is represented by human scholars, or anybody working in that area, and then of course there are questions about how Appalachia, or places like Appalachia that are outside the dominant narrative, are represented in these models. Do you have any thoughts on all your work in the field, regarding those discussions of representation that probably arose pre-generative AI, and where that conversation might be occurring now in terms of how the models are representing Appalachia? We’ve seen that as the models and the training datasets expand, you can add representations, but even adding that representation introduces some kind of value choices.

Dr. Theresa Burriss: Yes, there is. As you noted, for literally centuries, Appalachia has been marginalized and reduced to stereotypes. Even to this day, certain writers have not contributed in a positive way to that. I’m not going to name any names right now, but I think that there is a real danger in relying on it.

I actually have a case study that I can share now on the podcast, or that I’m going to share at the conference, about how a colleague of mine at a state institution was outraged at a conference. Some faculty members were teaching a community engagement class, and instead of engaging the community—going out and engaging Appalachians—they created Appalachians with ChatGPT.

And so consequently, given the web—because that’s where it pulls from—all of that negative stereotyping is going to come to the fore, and it’s not going to produce the very real, nuanced, diverse populations that exist in our region. So, I think that there are real dangers. And also, I mean, I’m sure you all have talked about the algorithms and how they multiply, and if it’s catching onto something, it’s going to keep pulling from that. It’s like a self-fulfilling prophecy, no matter what the group is. So, I have real concerns.

Subscribe now

Simulated Populations and Self-Representation

Anand: To ask you a little bit about that case study, I think what I find most fascinating about it is that this is a move that’s happening with lots of populations. There have been a lot of discussions, and I’ve heard some very optimistic views, on the idea of simulating environments and simulating populations within the use of generative AI to represent different groups. So maybe instead of polling actual people, you can run a simulated environment of a large population to be able to see how they’ll react to a polling item, or an election, or what they might do in a market economy. That point about the construction of that community or that population drawing from generally available materials—the assumption, of course, for many of us is, well, that’s just the way they’re already represented, so why wouldn’t we want to continue that representation? But there’s not a chance for that population to better represent themselves within that, and it further kind of entrenches this marginalization as you were talking about.

Is there a good way—aside, let’s say, that we can’t just stop those types of representation—is there a better way of going about that? Is there a better way to make sure that marginalized groups can own or control their own representation within generative AI, or within population studies, or within simulations? Obviously, in that case you mentioned, you make a good case that it shouldn’t have been done that way. But what if we’re worried that it’s going to continue to be done? How do we counter that?

Dr. Theresa Burriss: I think the only way that that can be countered—and oh my gosh, I mean, we’re talking years and years of information that has been accumulated on the web, and to undo that and to provide a counter-discourse. I think about counter-narratives; even still, even in print, that is a challenge.

And this is, I think, one of the dangers of AI: that it continues to, as you said, marginalize marginalized groups. I mean, who has access even to the internet? I mean, we think it’s a universal thing, but it ain’t. I mean, even where I live between Bristol and Abingdon, we did not get broadband until it was like six years ago. And there are still places over in the coalfields where I can go and I don’t have a cell phone signal.

So, I don’t think that there is an easy fix to that. It’s something that pollsters, researchers, educators, students, and politicians all need to be aware of. I mean, the paradigm that you’re operating from is not the paradigm of so many others.

Subscribe now

Data Centers, Environmental Justice, and Appalachia as Sacrifice Zone

Stefan: You know, it’s interesting. You talk about how some people are still disconnected, at least in terms of broadband, which you kind of need to do all but the simplest tasks, right? So, it’s an area that’s disconnected, but then they have trouble connecting all the areas, but they don’t have any trouble building the data centers, right? We’ve always had data centers, but Virginia has become kind of the most significant location, right, in the conversation globally.

I’m wondering, with your seat on the Virginia Council on Environmental Justice, there are a lot of issues surrounding data centers. What are some of the things you think about regarding environmental justice? I remember studying this decades ago for a debate topic and learning about brownfields and all those types of things. But how do you kind of contextualize the data center debate and conversation in terms of environmental justice?

Dr. Theresa Burriss: I think that it is another environmental justice (EJ) issue. Just as we have seen—whether it’s brownfields or, I think about the Kingston coal ash spill that occurred in northeast Tennessee, I think it was in 2008. The cleanup—all that coal ash went to a poor, Black community in Alabama. This is just another form of environmental injustice because typically—not always, but typically and historically—it’s “not in my backyard.” If you’ve got the money and the political access, it ain’t going to happen in your backyard.

It also feeds into how Appalachia historically has been known as a national sacrifice zone because of our coal. First, it was timber, then it was coal. And we’re still extracting coal—not just deep mining, but also surface mining. With the energy required and our reliance still on fossil fuels, it remains to be an issue.

I’ve taught environmental and social justice classes, and so oftentimes these communities are held hostage. Because it’s a single industry, they don’t have other opportunities to make the same kind of money. So, they’re held hostage. And it’s not like the industry or corporations can go someplace else, because the fossil fuels exist in particular places.

But I am very concerned not only about the environmental health but the physical health of the humans. I’ve done a lot of work on black lung disease with coal miners in central Appalachia. That was on the decline because of laws that were enacted under President Carter in the 70s, but it has skyrocketed again for a variety of reasons. But here again, it’s a major concern.

And we haven’t even talked about the water consumption that is necessary for the cooling of these data centers. I know that in the Virginia legislature, there are several bills that are addressing data centers from different capacities. For example, one is to require them to reveal the volume of water intake. Yeah, I mean, we have a long way to go to figure all this out.

I don’t think that AI is going to go away. I don’t think data centers are going to go away; I’m a realist in this regard. But one of my friends who is in a different industry but is very environmentally conscious sent me an article about the untapped—pun intended—opportunity with geothermal, and that could be something that is much more environmentally friendly, I’ll say.

Local Leadership and the Civic Stakes of Data Center Siting

Anand: So when we think about communities that are grappling with this, figuring out whether we welcome a data center in or not, it seems like most of the reaction has been one extreme or the other. Either we welcome them in and we cover everything up because we have a tax base there that’s going to benefit the community—I have a colleague who was doing FOIA (Freedom of Information Act) requests to get local boards of supervisors to release the information about how much water was going to be used, and they were really just trying to hide that—and then you have the other extreme where it’s just the NIMBY (”Not In My Backyard”) response: “We don’t want it here, you can’t do it here.”

Is there a case to be made that we need to be open to data centers coming into our communities because we have a better chance of regulating them, of making sure they use renewable energies, explore geothermal, have a way to clean water, recycle water, and have better cooling systems and processes? Getting back to tension, is it difficult to take an environmental justice position and welcome the data centers in with the idea that they’re going to go somewhere, and at least we can better monitor and regulate them? Or would it be better to say no, let them go and manage it somewhere else? How do we work through that in practical terms, even if we recognize that by going through that we are kind of infringing on some of the values that we have?

Dr. Theresa Burriss: I think it is dependent on municipality leadership. To your point about your colleague submitting FOIAs because the board of supervisors was withholding that critical information: if you were going to have a data center in your community, it is up to leadership to require and hold accountable all the information, and to mandate, “If you’re going to come here, you have to follow these policies and procedures,” instead of just turning a blind eye and saying, “Oh, it’s all about the...” Well, and they get big tax breaks too, which I don’t understand.

So yeah, I would say the bottom line is it depends on leadership, and that’s why it is so critical that the community votes. That’s where, no matter your race, class, or gender, go vote, because they are the ones who are going to be in charge if you have these data centers come in.

Subscribe now

The Liberal Arts: Decolonizing the Imagination

Stefan: We focus a lot on the liberal arts, especially with the center, and it’s kind of a broad conceptualization. Lots of times when I have these kinds of opportunities, I ask: what do you think of when you think of the liberal arts? What is a core identifying principle or idea, and how do you think that can help facilitate these discussions—whether they’re related to authenticity, the environment, the overlap between the two, or the contextualization of Appalachia? How do you think we could bring that forward in a humanities-focused education?

Dr. Theresa Burriss: I think of the liberal arts as an education in critical thinking, in creative thinking, and in what does it mean to be an engaged, informed, contributing citizen. It invites you—this is what I loved about being a philosophy major. Even though my mechanical engineering father asked me, “What the heck are you going to do with that when you graduate?” philosophy invited me to think about ideas I had never even thought about.

Interestingly, I just finished writing a keynote address that I’m giving tomorrow at Emory & Henry University for the Mosaic ceremony, which is a cording ceremony before graduation on Saturday for underrepresented groups. I use bell hooks as my launching pad, and how she talks about the importance of not only decolonizing our minds, but our imaginations. I feel like the liberal arts do that for us. They help guide us to think about how the very strong social messaging we get oftentimes is used to control, to further power, and to disenfranchise. So, I very much see the liberal arts playing a critical role in AI for all of those reasons.

Democratizing Voice: CityGov and the Non-Writer Expert

Anand: Thinking about the keynote that you’re giving and some of the discussion we’ve had about the way that AI can homogenize the voice—it can bring people more toward the median and remove a lot of the difference. Unfortunately, some of the arguments given for expanding data centers, expanding the technology, and expanding AI is that it democratizes access; that it no longer requires the same education, or access to education, or to means of technology. It’s cheaper, it’s more readily available for people to be able to write, translate, and express themselves. Back to the tension between the two: is there a way to reconcile the two? Is there a way to move forward? Is it about recognizing that tension?

Dr. Theresa Burriss: So, I’m involved in this open platform called CityGov. These are some friends of mine who live in New York, and two of my friends are going through a master’s program at Harvard. Sam came up with CityGov, and he told me, “I employ your language because of what you, Theresa, said about trying to reconcile.”

AI is a part of it. It doesn’t have to be for people who are contributing articles to this open platform, but what AI does—not everybody is a writer, and I recognize that as a writer, but there are people who have incredible experience and best practices that should be shared, whether it’s with other municipalities, educators, healthcare providers, or whatever the realm. You might not be a great writer, but you know what you’re doing and you’ve got a track record of proven success. Your democratized voice through AI makes what you know accessible to others.

I think about my two sons; they are very different. My older son is very much an existential nerd like his mother. My younger son—school was not for him. He works in the field in construction, in commercial construction. Both of them are brilliant. If Campbell, my younger son, wanted to share best practices, writing is not his thing. He hates it. But if AI were to enable him to share what he knows are best practices in the field—I mean, he educates me all the time just by talking to me—look at what he could do through AI. So, that’s how I see that platform.

Subscribe now

Navigating the Extremes: AI Refusers vs. Concentration of Power

Stefan: I was thinking just about a couple of the tensions. I always see—I’ve been telling people—well, if we don’t build more data centers, then the people who are going to get access to the AI are the biggest corporations and the military. It’s going to get really expensive, and you already kind of see some of that lately. If we don’t have more data centers, then you and me, we might lose meaningful access.

And then I think—I don’t know if you saw—but Anthropic just made a deal with Elon Musk to use all his data centers. I’ve really never used Grok. I tried it a couple of times, but I would never pay because I’m not an Elon Musk fan, and the other ones could do the same thing I needed it for. So, it really wasn’t much of a moral quandary. But now that line’s blurred. I love Claude, and I probably do use that all the time. So it’s like, well, what am I supposed to do now? Am I really just going to—I’ll keep using Claude. But there aren’t just so many clear answers to lots of these quandaries, right? It’s a general-purpose technology. It’s going to be used for things we like and we don’t like.

When you have conversations with people, sometimes they are very much, “I don’t want to use AI because it causes environmental harm.” How do you engage in those conversations when you meet extremists on one side or the other?

Dr. Theresa Burriss: I have both in my life. In fact, I think about my colleagues at Transilvania University in Brașov, Romania, where I did my Fulbright. They are in the Faculty of Letters, and they have a zero-tolerance policy for AI. Now, that doesn’t mean that Romania isn’t employing AI and creating data centers, because I just read a newsletter today talking about a new data center there, but that’s their position. I even have a colleague at Radford University who—do you remember blue books, everyone? Yeah, they’re going back to blue books.

There’s something to be said about that. But I feel like to put your head in the sand is not doing your students a favor. I’m not dismissive of my friends and colleagues who do that; I understand the principle behind it. But to think that AI is not going to be a part of our lives—I think that’s just not realistic. So Stefan, I don’t know. I think it’s just a tough conversation when you’re having it on both sides. I try to—I’m a negotiator and try to be a peacemaker and find that middle ground, and engage in authentic, meaningful discussion.

Closing: Hope, Biophilia, and the Pull of the Real

Anand: And that’s what we hope you do when you come to campus in July. That’ll be great. We’ll put you to work as the peacemaker and see how we can bridge some of those differences wherever possible. We’re so appreciative of you taking your time to talk with us.

The last question I just wanted to ask: what gives you hope, maybe in a Derridean or existential sense, and/or what’s the question you’ve not yet figured out how to answer?

Dr. Theresa Burriss: I think what gives me hope is that people are pushing back on this intense trend—going back to the conversation about corporations—we are in a for-profit society, but I think it is so important that people push back on that, and I see it happening. I don’t know that it’s happening at the pace that it should be happening.

The questions—there are so many questions, and there are questions I don’t even know are questions, and that’s part of life too. But my hope is that—I mean, I’m looking outside my window. I live in the most beautiful part of the world in the Appalachian Mountains, and I do not want people to substitute the artificial for the real.

I think about the natural environment, going back to the origins of place-based education. I think about E.O. Wilson’s biophilia hypothesis and that innate desire he says humans possess to be connected to nature. I believe in that, and I hope—and I see evidence of that—persisting in spite of the artificial intelligence proliferation, if you will.

Anand: Well, that’s great. Thank you for that. For anybody that’s watching that wants to learn more about your work, where can they find you other than on July 22nd and 23rd at UMW?

Dr. Theresa Burriss: They can just do a Google search, or they can do an AI search and find me.

Anand: Excellent. Excellent. Well, thanks so much for joining us. I look forward to welcoming you to campus, and enjoy that beautiful nature that you live in.

Dr. Theresa Burriss: I indeed will. Thank you.

Subscribe now

Stefan-Bauschard | GlobalAI Debates

I have something of a reputation as a provocateur. I’ll write a take partly because I think the conventional wisdom deserves a stress test, that too many humans have become stochastic parrots, or just because I think a new industrial era requires a new way of thinking. Simply adding AI to our current socioeconomic system is not what is unfolding.

This is not one of those takes.

I am writing this one straight, even though I realize it will not get a lot of support, as most Americans oppose data centers.

That said, the growing movement to slow, block, or impose a moratorium on data center construction, while with good intentions and reflective of important concerns, is one of the most quietly damaging policy positions in American life right now — and the people advancing it are about to support the rationing of superintelligence itself to the Pentagon and the largest companies on Earth. Block the buildout and you don’t stop superintelligence from arriving. You just guarantee it arrives wearing a uniform or a corporate logo, with no third tier of academic, civic, or public infrastructure for everyone else.

Let me explain.

Subscribe now

The fantasy of stopping AI

Before we go further: nobody is going to stop AI.

A lot of the policy conversation right now is conducted as if there is some version of the next ten years where the data centers don’t get built, and the models don’t get trained. There isn’t. The geopolitical incentives, the capital flows, and the national security logic all point forward.

If your county turns down a project, it goes to the next county. If your state turns it down, it goes to the next state. If county and state deployment slows down, the centers get built on federal lands, in the Gulf, in Asia, and — within a decade — in orbit. Microsoft has already prototyped subsea data centers. SpaceX is openly designing for compute in space. The Pentagon is not going to wait on a county zoning board.

The question on the table is not whether the largest AI systems in human history get built. They are getting built. The only live questions are where, for whom, and whether we build enough to benefit everyone.

A moratorium does not stop AI. It decides which communities get the tax base, which get the jobs, which get the leverage — and if all citizens get cheap, abundant access to the resulting superintelligence.

Subscribe now

What happens when you restrict compute

Compute is the new electricity. That isn’t a metaphor anymore; it is an operating fact of the economy.

Restrict data center construction and two things happen, both predictable. First, compute becomes scarce — and scarce compute does not flow to the people who need it most. It flows to the people who can pay the most: the Department of Defense, the major cloud hyperscalers, the largest financial institutions on the planet. The middle of the economy — the small business, the independent researcher, the regional hospital, the school district, the solo developer trying to build something — gets crowded out. Not because they don’t have good ideas. Because they can’t outbid Lockheed.

Second, the compute that remains gets more expensive. Restrictions raise the price. They always do. The hyperscalers absorb it. Everyone else doesn’t.

If you imagine the most concentrated, most corporate, least democratic possible future for AI — that is what restricting data centers produces. Not the opposite.

Subscribe now

The coal analogy

Imagine that in 1925 — worried about coal pollution, and rightly so — we had decided to limit power plant construction until cleaner alternatives matured. No new coal plants. Strict caps on the ones we had. Wait for nuclear. Wait for renewables.

Would electricity have stopped? No. It would have been rationed. The wealthy and the well-connected would have gotten it. Firms big enough to build their own generation on-site would have gotten it. Rural electrification — the single most economically equalizing infrastructure project of the 20th century — would not have happened. The middle class would not have been built.

The pollution argument was real. The answer was not to restrict the infrastructure. The answer was to clean it up while building it.

I am asking us to remember that lesson.

Subscribe now

Intelligence rationing

Here is where the stakes get higher than electricity, which is saying something.

The education system is already a scarcity machine. There are only so many seats at Stanford and only so many at Exeter, and we have built an entire social hierarchy around the competition for them — because getting one confers a lifetime advantage. We have made our peace with rationing prestige.

Now compound that with AI.

The richest universities and the most expensive private schools can pair entrepreneurship programs with access to the most capable frontier models. Their students will graduate having spent four years working alongside something close to a personal genius. Everyone else will graduate having spent four years working alongside whatever the free tier happened to allow that week when it wasn’t compute-constrained.

The alternative to building data centers is building intelligence rationing. It will look exactly like medical rationing looks today. The people with means get the good version. The people without means get the version that’s “good enough.” Outcomes diverge. Mobility collapses. The story we tell ourselves about merit gets even thinner than it already is.

Dario Amodei talks about putting “a country of geniuses in a data center.” Fine. I’ll take him at his word. The question is: who gets to talk to the geniuses?

If we build enough compute, the answer is everyone. A nurse in rural Ohio. A teacher in West Texas. A high school senior who has never met anyone who went to college. A small business owner in Detroit. Compute empowers.

If we don’t build enough compute, the answer is the people who can afford it. And we will have engineered, with our own good intentions, the most stratified society in modern American history.

Subscribe now

Stop fighting the centers. Start negotiating the terms.

If you accept the premise of this essay — that the data centers are getting built somewhere regardless of what your county does — then the leverage local communities have is not the leverage to prevent construction. It is the leverage to shape it. To decide, while there is still time to decide, what a host community gets in return.

That leverage is enormous. We should use it.

A tax base you cannot move. A data center is one of the few pieces of physical capital in the modern economy that cannot be relocated when the tax rate changes. You can move a factory. You cannot meaningfully move a 200-megawatt facility with billions of dollars of fiber, substations, and cooling infrastructure sunk into the ground. That makes data centers an unusually durable tax base — and we are going to need one. Nobody, including the people building these systems, actually knows what AI will do to employment. If the disruption is severe, communities will need a way to fund the response. Whether that response looks like a partial UBI, expanded community services, retraining, or something we haven’t named yet, you cannot fund it without a tax base. Lock it in now.

Commitments to schools and community. Make the deal include a real commitment to fund the local schools — not just generally, but specifically, including frontier AI access for every student in the district that hosts the facility. If the company is going to draw on your power grid, your water table, and your zoning code, your kids should be the first in the country to use what it makes.

Power they generate, prices the community pays. Data centers are among the largest power purchasers in America, which means they have an interest in building generation and the capital to do it at scale. Make them. And then make a piece of that generation flow back to the host community at subsidized rates — or, in the most ambitious version, free residential electricity within a defined radius. That isn’t radical. It is the same logic the railroads operated on a century ago and the rural electrification co-ops a generation after that. If we are going to live next to the largest energy consumers in the country, we should also live next to the lowest electricity bills.

Environmental terms that actually bind. Water, cooling, emissions, noise, heat island effects. Negotiate them. Inspect them. Penalize violations. None of this is the same as banning construction. It is the price of doing business, and a serious community can demand a serious price.

The point is to turn the data center buildout from a zero-sum fight — they win, the community loses — into a positive-sum deal. Compute gets built. The community gets a tax base, a power deal, a school commitment, and environmental guardrails. The country gets to remain the place where the future is built.

Subscribe now

The world we could choose

For the first time in human history, we are within reach of ending the scarcities that have defined modern life. Not luxury scarcity. Human scarcity. The scarcity of a good doctor. A patient tutor. A competent lawyer. A mentor for a kid who has never had one.

These have been scarce forever — not because the knowledge didn’t exist, but because the humans who held it could only be in one place at a time. Expertise has always been bottlenecked by the carrying capacity of a single human life.

AI breaks that bottleneck. A country of geniuses in a data center is also a country of the best doctors. A country of 1:1 tutors. A country of advocates for everyone who has never had one.

And here is what we have to be honest about: we already ration all of this. Health care and legal help are rationed by income and zip code. Education is rationed by property tax base. Mental health care is rationed so severely it functions, for most Americans, as something that doesn’t exist.

We made our peace with it because there was no alternative. You cannot manufacture more cardiologists overnight. The scarcity felt like physics.

It isn’t physics anymore. It’s a choice.

Build the compute, and the scarcity breaks — directionally, durably, the way electricity broke the scarcity of light. Don’t build it, and the scarcity holds. Worse, it compounds. Because now there will be a cardiologist, a tutor, an advocate in the machine. The best versions just won’t be available to everyone. They will be available to whoever can pay for the frontier tier while everyone else gets whatever the free tier happened to allow that week.

The rationing we have lived with for centuries gets replaced by a rationing that is faster, sharper, and more consequential — because the gap between the frontier model and the free tier will be larger than the gap between Harvard and another college, and it will compound every year.

That is the choice on the table. Not whether to ration intelligence. We already ration intelligence. The choice is whether to use this strange, narrow, once-in-a-civilization moment to end the rationing — or to lock it in for another century with better technology and worse equity.

Moratoriums on data centers are a vote to lock it in. For a world where superintelligence exists but most people can’t afford to talk to it. For a world that looks a lot like the one we have now — stratified by access, sorted by zip code, defended by people who already got theirs — except more consequential, because the thing being rationed is no longer prestige or comfort but cognition itself.

Subscribe now

We have done better than moratoriums before. Rural electrification did better. The interstate system did better. The public library did better. Each was a fight. Each was won by people who understood that infrastructure is destiny and that you cannot democratize what you refuse to build.

Build the data centers. Tax them hard. Make them fund the schools and lower the electric bills. But build them. Everywhere. Fast.

Because the alternative isn’t a world without superintelligence.

The alternative is a world where superintelligence exists, and your neighbor’s kid can’t reach it.

We have spent a century rationing the things that matter most. We finally have a chance to stop.

Let’s not waste it building a wall around the cure.

Subscribe now

TL;DR

As the school year ends for many, I see twelve emerging trends in AI that will both create profound change and risk intelligence deprivation unless we chart a new course.

Subscribe now

The Twelve Trends

As the school year ends for many, I see 12 emerging trends in AI that will both create profound change and risk intelligence deprivation.

(1) Multimodal interaction. Originally, GAI was just trained on text, and people interacted with it through Q (human) & A (the AI). Now, training is multimodal, and interaction is multimodal and seamless (Pichai). We saw this most recently with the Thinking Machines interactive model releases, but it’s also in Meta glasses, and it’s emerging in OpenAI’s voice reasoning models. In the not-too-distant future, most interaction will be through voice and video.

(2) Rapid, iterative improvements of the models (e.g., ChatGPT 5.4+, 5.5+ 5.6 (coming)+; Claude 4.6, 4.7, 4.8 ( 57.9% on Humanity's Last Exam with tools), Mythos; Gemini 3.0, 3.1+, 3.5 (Flash, Pro due soon)).

These improvements are based on expanded training data sets, expanded parameters, potentially 10 trillion in Claude Mythos, longer context windows, more sophisticated reasoning systems, improved multimodal capabilities.

[Link]. The capabilities seem to be doubling every 4.5 months, and the upper limits of Mythos are too difficult to ascertain, as they are limited, perhaps, more by token usage than raw ability. We should expect a “step-function jump” in forthcoming models from Anthropic and Gemini.

Models also have increasingly agentic architectures that allow models to plan, act, evaluate, and revise their own work so that they can complete work that would take humans a long time (16 hours+). [METR]

[In the newest test of an improved Mythos model, it was able to complete a hacking task that would have taken a team of human experts 20 hours]

The important point is not that any single model release will be decisive. The important point is that the curve keeps moving, and it is moving quickly. Each generation is not simply “a better chatbot.” It is a more capable cognitive partner: better at reading, writing, coding, researching, tutoring, designing, simulating, persuading, and coordinating complex tasks. And we are now seeing new advances monthly to bi-monthly.

And the model is only one layer of the change. Software built on top of the model can make its capabilities much more useful in practice. A raw model might be able to explain algebra, generate feedback on an essay, or help a student brainstorm an argument. But educational software can wrap that model in a structured learning environment: tracking a student’s progress, giving targeted practice, adjusting difficulty, remembering past mistakes, aligning tasks to a curriculum, and giving teachers dashboards that show where students are struggling.

Harnesses (think Claude Cowork, Claude Code, Codex) expand this even further, especially when they connect the model to tools. A model by itself can reason about a research question. But a model inside a tool-using harness can search databases, open PDFs, compare sources, run calculations, generate charts, draft a memo, check citations, revise the output, and then produce a polished final product. In other words, the harness turns the model from a conversational assistant into something closer to an action system.

See also: Anthropic launches Claude for Small Business to embed AI into the tools you forgot you pay for | Anthropic expands legal AI offerings with new Claude Cowork plugins

People are also beginning to use models in combination with one another, not just as standalone tools. One model might be stronger at writing, another at coding, another at long-context analysis, another at mathematical reasoning, and another at image or video generation. A person, or an automated system, can route different parts of a task to different models, then use one model to critique, verify, or improve the work of another. This creates a kind of model ecosystem, where the value comes not only from the best individual model, but from the way multiple models are orchestrated together to take advantage of their strengths and catch each other’s weaknesses.

Subscribe now

(3) Recursive self-improvement (RSI). RSI means an AI system helps improve the next version of itself, or the tools, code, data, evaluations, training process, and research workflows that make future AI systems better. The strongest version of RSI would be a system that can be told, “make a better version of yourself,” and then independently redesign, train, test, and deploy a more capable successor. We are probably not fully there yet. But we are seeing partial forms of the loop: AI systems helping write code, run experiments, evaluate outputs, optimize algorithms, generate training data, and assist the researchers building the next systems. Anthropic’s new research agenda explicitly raises the question of how humans would govern AI systems if they are being used to “autonomously develop and improve themselves,” and it also calls for “fire drill” scenarios for a possible intelligence explosion. (Anthropic)

There are already recent examples that are arguably RSI-adjacent, even if they are not full recursive self-improvement. Google DeepMind’s AlphaEvolve is a Gemini-powered coding agent that evolves algorithms by making direct changes to code, testing variants, and selecting improvements. DeepMind describes it as a system that combines large language models with automated evaluators to design advanced algorithms, and its technical paper says AlphaEvolve “orchestrates an autonomous pipeline of LLMs” whose task is to improve algorithms by modifying code. (Google DeepMind) Sakana AI’s AI Scientist is another example: it can move through much of the machine learning research lifecycle, including generating ideas, writing code, running experiments, analyzing results, producing figures, and drafting papers. Sakana reported that AI Scientist-v2 produced a fully AI-generated paper that passed peer review at a workshop associated with a major AI conference. (sakana.ai) Nature’s coverage framed the same case more cautiously, noting both the significance and the limitations of autonomous research tools. (Nature)

The deeper concern is that these systems may close more and more of the improvement loop. Anthropic co-founder Jack Clark recently argued that there is now a meaningful chance that, by the end of 2028, an AI system could autonomously build a better version of itself.

This is largely consistent with AI-2027’s original project for RSI (June 2027). xAI cofounder expects it in 12 months and Sam Altman wants an automated AI researcher by 2028. Anthropic’s Jesse Mu is looking forward to RSI so he can go home and knit sweaters.

See also: AI startup Recursive emerges from stealth with $650 million to build self-improving AI.

But even without a fully autonomous “AI builds the next AI” loop, the speed of software development has changed. Engineers can now use AI to draft code, debug errors, write tests, refactor systems, generate documentation, review pull requests, and experiment with multiple architectures quickly. That matters because AI progress is not only about bigger models. It is also about faster iteration. When AI helps build the software, evaluation harnesses, data pipelines, agent frameworks, and research tools that support future AI, the development cycle compresses. The result is a softer but still powerful version of recursive improvement: humans remain in the loop, but the loop itself moves faster because AI is helping build the next layer of AI capability.

Subscribe now

(4) The same acceleration is now visible in mathematics and scientific research. For years, math was treated as one of the hardest tests for AI because real mathematics requires more than fluent language. It requires abstraction, multi-step reasoning, proof discipline, creativity, and the ability to avoid one tiny mistake that ruins the whole argument. That barrier is starting to weaken. In 2024, Google DeepMind’s AlphaProof and AlphaGeometry 2 solved four of six International Mathematical Olympiad problems, earning a silver-medal-level score. Nature described this as a landmark because the system solved three algebra and number theory problems and one geometry problem, including the hardest 2024 IMO problem, which only five human contestants solved. (Nature)

The pace then jumped again. In 2025, an advanced version of Gemini Deep Think solved five of six IMO problems, earned 35 out of 42 points, and reached gold-medal level performance. Just as important, Google DeepMind says this version worked directly from the official natural-language problem statements and finished within the 4.5-hour competition time limit, instead of requiring humans to translate the problems into formal code first. (Google DeepMind) Quanta called this a tipping point: mathematicians who had dismissed AI as too error-prone began finding that models could help “break genuinely new ground,” with Terence Tao saying that 2025 was the year AI started being useful for many different mathematical tasks. (Quanta Magazine)

The research-level progress is even more important than the contest results. Epoch AI’s FrontierMath benchmark was created to test advanced mathematics, including unpublished problems written and peer-reviewed by expert mathematicians and open research problems that remain unsolved by humans. (Epoch AI) Its Tier 4 problems are research-level, not ordinary school math. Epoch describes Tiers 1 through 3 as spanning undergraduate through early postdoc difficulty, while Tier 4 is explicitly research-level mathematics. (Epoch AI) In May 2026, Google’s AI co-mathematician paper reported that its agentic math workbench scored 48% on FrontierMath Tier 4, a new high among evaluated AI systems. The paper says the system solved 23 of 48 private Tier 4 problems, including three problems that no previously evaluated system had solved. (arXiv)

There are also striking examples of AI contributing to actual open problems. Terence Tao wrote about Erdős problem #1026, explaining that it was solved through a combination of existing literature, online collaboration, and AI tools. That wording matters. This was not a clean story of “AI solved it alone,” but it is still a major signal: AI is becoming part of the mathematical discovery process, especially when paired with human judgment, community review, and careful proof writing. (What’s new) Scientific American also reported that a 23-year-old amateur with no advanced math training used ChatGPT to crack a 60-year-old problem that world-class mathematicians had tried and failed to solve, using a method experts described as one no human had thought of. (Scientific American)

Most recently, planar unit distance problem, or Erdős problem 90, which “has intrigued mathematicians for decades.” (Melissa Lee). Most recently (May 28), another major problem, “this time in additive combinatorics, has fallen, this time to humans rather than AI, but using methods related to the AI solution to the unit distance conjecture.”

Most recently, Google DeepMind’s new paper shows that AI can now search for formal mathematics proofs, but only inside carefully constrained worlds. The striking result is not that the system “thinks like a mathematician,” but that it keeps forcing its thoughts through Lean, a proof system where every logical step has to compile. That matters because LLMs can sound convincing in math while still making small, fatal mistakes.

AlphaProof Nexus changes the role of the model from a persuasive storyteller into a generator of proof candidates that can be tested, rejected, repaired, and improved. The system lets an LLM keep editing a formal proof, read compiler errors, try again, and sometimes call on stronger proof tools for smaller subproblems.

The stronger version also maintains a shared pool of partial proof attempts, rates which ones seem promising, and uses them to guide later searches. The verifier is not a cosmetic add-on. It is the mechanism that makes exploration tolerable. Without it, a beautiful proof sketch can hide a false lemma; with it, the model has to turn insight into executable logic or fail visibly.

The results are significant: the best agent solved 9 of 353 formalized Erdős problems and proved 44 of 492 open conjectures from the Online Encyclopedia of Integer Sequences, while also helping with problems in optimization, graph theory, algebraic geometry, and quantum optics. But the failures are just as revealing as the wins, because the agents sometimes buried the hard part inside a helper lemma or hallucinated a known result, exactly the kind of error formal checking is designed to expose. The real shift is not full mathematical autonomy. It is a new division of labor: humans choose the formal question, libraries define the terrain, models propose routes, and the proof assistant refuses to be impressed.

The deeper point is that math is not just one academic subject. Math is the language underneath physics, computer science, cryptography, optimization, engineering, logistics, economics, materials science, and much of modern biology. When AI gets better at math, it gets better at discovering algorithms, modeling systems, designing experiments, proving constraints, and finding solutions in enormous search spaces. Google DeepMind’s AlphaEvolve shows this clearly: it discovered an improved algorithm for multiplying 4x4 complex-valued matrices, beating the best-known approach in that setting from Strassen’s 1969 work, and it also improved results in open problems across areas like geometry, combinatorics, and number theory. (Google DeepMind)

That is why the Solve Everything frame is provocative but useful. Diamandis and Wissner-Gross argue that the core bottleneck is no longer just money, labor, or even information. The bottleneck is scarce expert attention, and AI lowers the cost of applying cognition to hard problems. Their model emphasizes a full stack: goals, task taxonomies, observability, harnesses, model layers, actuation, verification, governance, and distribution. In other words, solving problems requires more than a smart model. It requires systems that can define the target, test progress, act in the world, and verify results. Their broader claim is that superintelligence should be aimed at problems that make life “short, expensive, or unfair,” with the “Intelligence Revolution” attacking the final bottleneck of scarce expert attention. (theinnermostloop.substack.com)

The Tokyo Institute of Sciences has opened the first fully automated medical lab.

Subscribe now

(5) We are also officially past the idea that AI is just a bubble. That does not mean every AI company will survive. It does not mean every valuation is rational. It does not mean investors will avoid losses. Many companies may fail, and some current spending will almost certainly be wasted. But that is different from saying AI itself is a passing fad. At this point, AI is too deeply embedded in corporate strategy, national security planning, infrastructure investment, software development, and government modernization to simply disappear.

One signal is the scale of real demand. Anthropic reported in February 2026 that its run-rate revenue had reached $14 billion, growing more than 10x annually over each of the previous three years. By April, Bloomberg reported that Anthropic’s revenue run-rate had topped $30 billion, up from $9 billion at the end of 2025. By May, Forbes cited $44 billion ARR, and the Wall Street Journal reported that Anthropic’s annual revenue run-rate was expected to hit $50 billion by the next month.

The trajectory is hard to dismiss: enterprise customers are paying for frontier AI at a scale that has moved far beyond experimentation. (Bloomberg). And their valuation — now $900 billion — has tripled since January. It just (May 28th) raised another $65 billion. If it hits $100 billion in ARR by the end of 2026, it will be worth $1 trillion (Blundin).

OpenAI hit $5.7 billion in a single quarter and has now surpassed 905 million weekly users.

NVIDIA just became the first company to crack a $5.5 trillion market cap.

Another signal is that the main bottleneck is no longer “does anyone want this?” The bottleneck is compute. AI companies, cloud providers, and governments are racing to secure chips, data centers, electricity, cooling, and grid access. [EpochAI]

Sam Altman may start a new AI compute company. Dell’s server revenue is up 757% to $16 billion.

The simple reality is that there isn’t enough compute for the thousands of agents people are spinning up (Blundin) or all of the new use cases people are imagining (Diamandis)

CNAS describes AI chip production as a “binding constraint” on the pace of the AI compute buildout, while CSIS argues that electricity supply may be the most acutely binding constraint on expanded U.S. computational capacity. In other words, the limiting factor is not imagination. It is physical infrastructure. (CNAS)

The corporate adoption story is also shifting from experiments to deployment. Anthropic’s business growth is not just coming from consumer curiosity. Axios reported that, in April 2026, Anthropic surpassed OpenAI in workplace AI adoption among businesses tracked by Ramp, with Anthropic at 34.4% adoption and OpenAI at 32.3%. Business Insider similarly described the shift as a major enterprise reversal, driven especially by Claude Code and the spread of AI into software development, legal, finance, and research workflows. (Axios)

The military dimension makes the “AI is going away” argument even weaker. The 2026 Department of War AI strategy says that military AI is a race and that “speed wins,” directing the department to measure deployment velocity and operational cycle time as decisive variables in the AI era. Reuters also reported that the Pentagon has moved to adopt Palantir’s Maven AI system as a core military command-and-control platform, capable of analyzing data from satellites, drones, radar, and sensors. Once AI becomes central to intelligence analysis, logistics, autonomous systems, cyber operations, targeting support, and battlefield decision-making, it becomes a strategic necessity, not a consumer trend. (U.S. Department of War)

Governments are also moving from policy statements to operational transformation. The UAE announced a plan to deploy agentic AI across 50% of government sectors and services within two years, using autonomous AI systems to transform public-sector operations. Whether that target is fully achieved is a separate question. The important point is that governments are now openly planning around AI as administrative infrastructure, not merely as a productivity app. (Emirates 24|7)

Jamie Dimon’s 2026 shareholder letter captures the same reality from the perspective of one of the world’s most important financial institutions. Dimon pointed to a huge increase in AI-driven capital spending and construction by the five hyperscalers, estimating about $450 billion in 2025 and approximately $725 billion in 2026. He also frames AI as productivity-enhancing and transformational, while acknowledging that the capital spending boom may create short-term inflationary pressure. That is the right balance: some money will be wasted, some firms will implode, some jobs will be displaced, and some early claims will be exaggerated. But the underlying technology is real. (JPMorgan Chase)

A quarter of all lobbyists are now working on AI issues. Industry is getting a say while many students are still being told that AI is a bubble. Companies are pressing their own interests, whereas in Korea and Poland they are already talking about economic redistribution.

Subscribe now

(6) Emerging consensus on what students need to learn.

I probably should not say there is a “consensus” on anything related to AI and education. That may be too strong. The field is moving too fast, the politics are too charged, and schools are still trying to figure out what problem they are actually solving. But there does seem to be an emerging center of gravity.

A lot of the debate so far has been about how students should be allowed to use AI in school: green light, yellow light, red light; what counts as cheating; whether teachers should ban it, limit it, detect it, or redesign assignments around it. Those questions matter, but they are not the starting point. The primary question has to be: What do students need to learn in order to live, work, think, and lead in a world shaped by increasingly powerful AI? Only after that do we ask: How can AI help students prepare for that world?

Going the other way makes no sense. If we start with tool-use rules, we end up protecting the current curriculum from disruption. If we start with the future students are actually entering, the conversation changes. The question is no longer, “How do we preserve the paper, the worksheet, the test, or the essay?” The question becomes, “What human capacities matter most when machines can increasingly read, write, code, reason, design, simulate, tutor, and advise?”

I would make three claims.

First, students need to understand the civilization-transforming developments and implications of AI. This cannot be treated as a side topic, a tech elective, or a one-day digital citizenship lesson. UNESCO’s AI Competency Framework for Students argues that students need to become responsible users and co-creators of AI, with competencies across a human-centered mindset, ethics, AI techniques and applications, and AI system design. (UNESCO) The OECD similarly argues that education systems need to understand AI capabilities and compare them with human skills, because that is how we begin to see which tasks AI may automate, where humans may be complemented, and how the demand for skills may shift. (OECD)

Trying to hide this from students protects the status quo curriculum for a little while. It lets adults carry on as “normal.” But it deprives students of important choices about their own education. They need to understand what is happening to work, knowledge, creativity, media, politics, warfare, science, and human agency. They need to know why AI matters, where it is overhyped, where it is already real, and where it may transform the assumptions underneath their future lives. We do not have to terrify them. But we do have to respect them enough to tell them the truth.

Second, students need much stronger soft skills, communication skills, and human interaction skills. I know “soft skills” is an imperfect phrase, but the point is becoming painfully obvious. The World Economic Forum’s Future of Jobs Report 2025 identifies AI and big data as the fastest-growing skill area, but it also highlights rising demand for creative thinking, resilience, flexibility, agility, curiosity, lifelong learning, leadership, and social influence. (World Economic Forum)

See also — Open to Work: How to Get Ahead in the Age of AI

This is why the continued obsession with paper writing as the center of academic seriousness feels increasingly painful. Writing still matters. But if AI can already write better than many students, and often better than many adults, then continuing to treat the traditional paper as the supreme evidence of learning is a problem. Meanwhile, many students are underdeveloped in the skills that AI does not automatically replace: speaking clearly, listening carefully, asking better questions, reading a room, resolving conflict, persuading real people, collaborating under pressure, exercising judgment, and building trust.

In an AI-saturated world, communication is not ornamental. It is infrastructure. Students will need to explain ideas, challenge bad outputs, work across teams, deliberate about values, and make decisions with other people. The irony is that schools often call these “soft” skills while designing schedules, assessments, and graduation requirements around the things machines are getting better at doing.

The American Association of Colleges and Universities regularly publishes a national survey examining how employers view the preparation liberal arts institutions provide for graduates entering the modern workforce. The most recent survey was completed in August 2025. They need the skills I’ve been discussing, and schools continue to under-invest in these relative to writing

Third, students need to learn how to use and interact with AIs intelligently. This does not mean every student needs to master today’s specific tools. The tools will change. The interfaces will change. The names will change. But the idea of working with silicon intelligence will not change.

The deeper goal is not “prompt engineering” as a trendy skill. It is helping students understand intelligence itself, both human and machine. What are models good at? Where do they fail? What does it mean to reason, predict, infer, simulate, generate, verify, or judge? How is human intelligence different from machine intelligence? How do you collaborate with a system that can produce fluent answers but still be wrong? How do you use AI without outsourcing your own thinking? How do you check its work? How do you ask better questions? How do you combine AI assistance with human purpose, ethics, taste, and responsibility?

That is the real curriculum shift. Students should not merely be protected from AI, and they should not merely be allowed to use it as a shortcut. They should be taught to work with it, question it, challenge it, direct it, and understand its limits. The goal is not AI dependence. The goal is AI fluency plus human agency.

This will facilitate their development as entrepreneurs, something many see as the job of the future. From Megatrends this week —

So, yes, there may not be a formal consensus yet. But the emerging direction is pretty clear: students need to understand the world AI is creating, develop the human capacities that become more valuable in that world, and learn how to interact with intelligent machines responsibly and effectively. Everything else, including classroom AI-use policies, should flow from that.

(7) More thinking about AGI. AGI is often discussed alongside superintelligence, and in some public conversations the terms are used almost interchangeably. More precisely, AGI refers to artificial general intelligence, meaning AI systems that can perform a wide range of intellectual tasks at or near human-level. Superintelligence usually refers to systems that go beyond human-level intelligence across most important domains.

The reason the two ideas are so closely connected is that once AI reaches general human-level intelligence, it may not remain there for long. Unlike humans, AI systems can potentially be copied, connected, accelerated, improved through vast amounts of data and compute, and used to help design even better AI systems. That creates the possibility of rapid recursive improvement: intelligent systems helping build more intelligent systems. In that scenario, AGI is not the endpoint. It is the threshold after which superintelligence could quickly follow.

This is not the end-all, be-all of the argument. AGI is not a single magical moment where everything changes overnight. It will probably be uneven, contested, and strangely gradual in some places while suddenly disruptive in others. Some people will say, “it is already here.” Others will say “it still does not count.” And because the definition of AGI is itself debated, we should be careful about treating any one forecast as settled truth.

But we also should not hide behind definitional uncertainty. One of the most reliable voices in the field, Google DeepMind CEO Demis Hassabis, has spent years offering one of the more careful and thorough definitions of AGI: not just a chatbot that sounds smart, but a general system that can understand, reason, plan, learn in real time, and help solve problems across domains.

And his timeline has moved. Around six months ago, Hassabis was still saying AGI was likely five to ten years away, and that one or two additional breakthroughs were still needed. (The Times of India) More recently, he has said we are roughly three -quarters of the way to AGI and continues to point toward something like a 2030 horizon. (GLN-7.5) That does not prove AGI arrives in 2030. It does mean that one of the people most responsible for building this technology now believes we are incredibly close by the standards of education planning.

Most recently, in an interview with Axios after his Google I/O keynote, he says that while he still expects AGI in 2030, 2029 is possible, with the next wave of agents as a “practice run” for AGI. Back in the 1990s, Ray Kurzweil predicted 2029 as the year we’d achieve AGI.

That is the part schools cannot ignore. A freshman entering high school today will graduate in 2030. I recently saw a student wearing an “X Class of 2030” shirt, and I could not stop thinking about it. She may be perfectly on track for the pre-AI world. She may be doing the assignments, earning the grades, preparing for college, building the résumé, and following the path adults have told her still matters. But will she be ready for a world of AGI, or near-AGI, or powerful agentic AI systems that can write, code, tutor, research, simulate, persuade, design, and increasingly act?

And it is worth noting that AI is already intelligent by traditional measures, exceeding average human intelligence on IQ tests.

But IQ does not measure “all of intelligence.” It measures a narrower but important bundle of cognitive abilities: how well someone handles abstract problems, patterns, symbols, language, memory, and speed under standardized conditions.

Modern tests like the WAIS and WISC usually break intelligence into areas such as verbal comprehension, visual-spatial reasoning, fluid reasoning, working memory, and processing speed. The WAIS-5, for example, explicitly expands coverage of verbal comprehension, visual-spatial ability, fluid reasoning, working memory, and processing speed. The APA defines an intelligence test as a test used to determine a person’s intelligence level by measuring abilities such as solving problems (APA Intelligence Test).

Schools have spent decades rewarding the kinds of cognitive tasks IQ tests measure: verbal reasoning, pattern recognition, memory, speed, and symbolic problem-solving. AI is already getting very good at those. The scarce human capacities are shifting toward judgment, communication, ethical reasoning, collaboration, taste, agency, and the ability to ask better questions.

Subscribe now

(8) Robotic capabilities have expanded sharply in just the last three months. Not in the “Rosie from The Jetsons is ready for every home tomorrow” sense. That would still be too strong. However, the direction is unmistakable: robots are becoming increasingly adept at navigating human environments, handling complex physical objects, coordinating with other robots, learning from foundational models, and transitioning from demonstrations to industrial deployment. Robots are giving a body to AI.

A new company, Shift, is even willing to clean your apartment in NYC for free in exchange for the right to record everything the cleaner does in order to boost training robots for home environments. Waymo is expanding the size of its robotaxis.

The biggest shift is that robotics is becoming part of the broader AI acceleration. For years, robots were mostly impressive in controlled environments: warehouses, factories, labs, or carefully staged demos. The new frontier is physical AI, meaning AI systems that do not just generate text, images, or code, but perceive the world, reason about objects, and take physical action. NVIDIA’s March 2026 announcement of Cosmos 3, which it describes as a world foundation model combining synthetic world generation, vision reasoning, and action simulation, shows where the field is going: robots can increasingly be trained in simulated worlds before being deployed in real ones. (NVIDIA Newsroom)

One of the clearest examples is Figure AI. In May 2026, Figure showed two Helix-02-powered humanoids autonomously resetting a bedroom in under two minutes. They opened doors, hung up clothing, put away headphones, closed a book, took out trash, pushed in a chair, and worked together to make a bed. The important part is not just that they completed a list of chores. It is that the robots were operating with a single learned vision-language-action policy in a messy home-like environment, including deformable objects like bedding and coordination between two robots. (FigureAI)

That is a big deal because homes are hard. A factory task can be structured. A bedroom is chaos: soft blankets, changing object positions, doors, furniture, trash, clothes, and unpredictable arrangements. Making a bed sounds trivial until you remember that a blanket is not a rigid object. It folds, bunches, slips, and changes shape. A robot that can handle that is moving closer to general-purpose physical labor.

This is the robotics version of the AI model shift, and Figure did run a demo of the robots 8, and then 24, hour shifts. Figure founder Brad Adcock predicts full general- purpose robots within a year.

Just as language models become more useful as they move from narrow classifiers to general-purpose systems, robots become much more powerful when they are trained on broad mixtures of human video, robot trajectories, simulation data, and synthetic data. NVIDIA describes GR00T N1 as trained on diverse data, including egocentric human videos, real and simulated robot trajectories, and synthetic data. (NVIDIA)

The third shift is that humanoids are moving from lab demos into actual industrial planning. Reuters reported on May 13, 2026, that British robotics company Humanoid plans to deploy 1,000 to 2,000 humanoid robots at Schaeffler’s global manufacturing sites by 2032, beginning with deployments in Germany from late 2026 into 2027. The initial uses include box handling and near-full-scale production-line testing. (Reuters)

China’s robotics push also accelerated visibly. Unitree’s May 2026 GD01 “manned transforming mecha” is partly spectacle, and we should not confuse it with a useful household robot. But even spectacle tells us something about hardware progress. Unitree demonstrated a 500 kg machine that can switch between bipedal and quadrupedal modes and is being marketed as production-ready, even if practical applications remain unclear. (The Verge) More broadly, China is using its manufacturing base, component supply chains, and policy support to push embodied AI and humanoid robotics forward quickly. (eWeek)

The projections are now enormous, but they should be read as directional signals, not guarantees. Goldman Sachs projects the humanoid robot market could reach $38 billion by 2035, with shipments rising to about 1.4 million units, a major upward revision from an earlier $6 billion estimate. (Goldman Sachs) Morgan Stanley is far more aggressive, estimating that the humanoid market could surpass $5 trillion by 2050, with more than 1 billion humanoids in use, mostly in industrial and commercial settings. (Morgan Stanley)

Subscribe now

(9) But even if intelligence is exploding, the real world still has constraints. AI may feel digital, weightless, and infinitely scalable, but it is not. Intelligence runs on hardware. Hardware sits in data centers. Data centers require land, chips, transformers, cooling systems, water, transmission lines, power plants, grid interconnections, permits, construction crews, and time. So even if the models keep improving quickly, the deployment of intelligence into the world will be shaped by very physical bottlenecks.

The most immediate constraint is compute, and compute is increasingly constrained by energy. The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global electricity use, and that demand has been growing at about 12% per year over the last five years. The IEA also notes that this is no longer a marginal issue for the power system because AI data centers require large, reliable, high-density electricity supply. (IEA)

That constraint is already showing up in concrete fights over infrastructure. The Guardian reported that data centers now use about 6% of total electricity supply in the U.K. and U.S., according to research by the International Data Center Authority, and that U.K. data center power use may quadruple by 2030. The same report notes that grid-connection delays are rising, which means the limiting factor is not simply whether companies want more AI. It is whether the power system can physically support it. (The Guardian)

This is why the AI buildout increasingly looks less like a software race and more like an industrial race. Brookings notes that one estimate puts global data center energy consumption near 1,050 TWh by 2026, which would make data centers, if they were a country, one of the largest electricity consumers in the world. (Brookings) Harvard’s Belfer Center similarly emphasizes the uncertainty but also the scale of the challenge, noting that projections for U.S. data center electricity demand by 2030 range from about 200 TWh to over 1,000 TWh. (Belfer Center)

So there is a strange tension here. On one level, AI is becoming more capable incredibly quickly. On another level, the next stage of AI progress depends on old-fashioned infrastructure: power plants, substations, transmission lines, water rights, zoning fights, and supply chains. A proposed Utah data center illustrates the scale. The Guardian reported that the Stratos project could require about 9 gigawatts of power, more than the entire state’s current consumption, triggering backlash over energy, water, pollution, and the Great Salt Lake. (The Guardian)

Robotics faces a parallel constraint. Embodied AI does not just need models. It needs bodies. If AI is going to move from screens into factories, warehouses, hospitals, farms, homes, and military systems, someone has to build the robots. That means motors, actuators, sensors, batteries, chips, cameras, gearboxes, rare earth magnets, carbon fiber, lithium, and enormous manufacturing capacity. McKinsey argues that humanoid robotics’ hidden constraint may be the supply chain itself, especially as companies try to scale actuators, sensors, batteries, and precision components. (McKinsey & Company)

Rare earth minerals are a particularly important bottleneck. S&P Global reported that humanoid robots could strain rare earth supply chains because each robot may require 30 or more motors, many of which depend on rare earth magnets for high power density. (S&P Global) A May 2026 study summarized by EurekAlert projects that production of humanoid and quadruped robots could grow by up to 100x by the late 2030s, creating new demand for rare earth metals and carbon fiber that could pose supply problems. (EurekAlert!)

Some people may celebrate these constraints. They may say, “Good, that means AI will not move as fast as the hype suggests.” And in a narrow sense, that may be true. Compute, energy, data center construction, robot manufacturing, rare earth processing, and grid capacity may slow deployment. They may create friction. They may prevent the most extreme overnight scenarios.

But that is not necessarily comforting. Scarcity does not make AI harmless. Scarcity makes AI expensive. And if the best compute becomes more expensive, access will not be distributed evenly. The most powerful AI systems will be available first to the wealthiest companies, the largest cloud providers, the most capitalized labs, and the military. The constraint will not stop the AI race. It may simply make the race more unequal.

That matters for education. If intelligence becomes abundant only for the powerful, then AI does not automatically democratize opportunity. It can widen the gap between students, schools, companies, and countries that have access to frontier tools and those that do not. Wealthy institutions will get stronger tutors, stronger research systems, stronger agents, stronger robotics, and stronger decision-support tools. Everyone else may get the cheaper, weaker, slower, more constrained version.

So the real-world constraints cut both ways. They may slow the timeline. They may make the transition messier. They may prevent a perfectly smooth explosion of intelligence. But they also concentrate power. In a world where compute, energy, and robotics supply chains become strategic resources, the question is not just whether AI gets smarter. It is who gets access to the smartest systems, who controls the infrastructure, and who is left using yesterday’s intelligence while the most powerful actors move ahead.

Subscribe now

(10). Widening Inequality and intelligence deprivation. When chatbots were first released, they felt surprisingly democratic. The basic version was free. The better version was often around $20 per month. That wasn’t nothing, especially for lower-income families, but it was still within reach for many people. For a brief moment, it looked like advanced AI might become a broadly available public utility: a tutor, writing coach, research assistant, coding partner, translator, brainstormer, and intellectual companion available to almost anyone with an internet connection.

That moment is fading.

The frontier AI world is becoming tiered. OpenAI now describes ChatGPT Plus as the $20 plan for lighter use, while its Pro tier is designed for higher-stakes work with significantly higher usage allowances for tools like Deep Research and Codex. (OpenAI Help Center) Anthropic moved in the same direction with Claude Max, offering higher-usage plans at $100 and $200 per month, far above what many students and families can afford. (The Verge) Google’s AI Ultra plan is even higher, priced at $249.99 per month and marketed as unlocking the “highest level of access” to Google AI, including higher limits and first access to some advanced models and tools. (Google One)

And that is just the consumer layer. Once you move into API access, coding agents, deep research tools, enterprise deployments, long-context workflows, video generation, and high-volume usage, the costs can rise quickly. The result is not simply “some people pay for premium features.” It is the emergence of an intelligence market, where the best reasoning, highest usage limits, fastest access, deepest research, strongest coding tools, and most capable models may increasingly be available first to wealthy individuals, elite schools, major corporations, governments, and militaries.

That creates something more serious than a digital divide. It creates a risk of societal intelligence deprivation.

We have often divided the world by access to land, capital, schooling, race, gender, class, language, geography, and credentials. Higher education has also created a kind of forced intellectual scarcity: access to elite professors, research networks, internships, admissions signals, and high-status credentials is rationed through tuition, geography, sorting systems, and institutional prestige. Now we may be building another layer of inequality on top of all of that: a division based on access to silicon intelligence.

UNESCO already describes this as an emerging AI divide, meaning unequal access to the benefits and opportunities created by AI across regions, communities, and socioeconomic groups. UNESCO specifically warns that marginalized communities bear the brunt of this divide and argues that AI literacy is essential to closing it. (UNESCO) That language matters because the issue is not only whether students can open a chatbot. It is whether they can access the most capable systems, understand how to use them, trust themselves around them, and build the human skills needed to direct them.

This is where the school-device issue becomes especially troubling. Many students do not have equal access to laptops, reliable broadband, paid subscriptions, or private workspaces at home. Pew reports that 93% of teens in households earning $75,000 or more have access to a desktop or laptop at home, compared with 78% of teens in households under $30,000. (Pew Research Center) Pew also reports that low-income adults remain far less likely to have home broadband, with a major gap between households under $30,000 and high-income households. (Pew Research Center) So when schools block AI tools on school devices, they are often blocking access on the only reliable devices some students have.

That is the cruel irony. The students who most need school-based access to AI may be the students most likely to lose it. Wealthy students can use AI at home, on personal laptops, with paid accounts, private tutors, and parents who know how to help them navigate the tools. Lower-SES students may depend on school-issued Chromebooks, school Wi-Fi, and school permission. If those devices block AI, then the school has not protected them from harm. It has protected them from opportunity.

The current anti-ed-tech reaction makes this worse. There are real reasons to restrict phones and devices during parts of the school day: distraction, social media addiction, bullying, cheating, sleep, attention, and mental health. Those concerns are real. But the response is increasingly blunt.

The problem is that the anti-device mood can easily slide into an anti-AI, anti-ed-tech, anti-access posture. Schools may think they are restoring focus by ripping devices out of students’ hands. Sometimes they may be. But if they do this without building alternative, supervised, equitable access to AI, they risk locking an entire generation of lower-income students out of the next industrial revolution.

That is not an exaggeration. AI fluency is becoming a form of economic preparation. Students who learn to use AI well will have access to better tutoring, faster research, stronger writing support, more coding help, better feedback, more creative tools, and a deeper understanding of how work is changing. Students who are kept away from AI will not become more “authentic.” They may simply become less prepared.

The danger is that schools will create a two-tier world. Wealthy students will use AI privately, strategically, and increasingly powerfully. Lower-income students will be told to put the device away, write the paper by hand, and wait for permission. The privileged will learn to collaborate with artificial intelligence. The less privileged will be protected from it.

That is societal intelligence deprivation. Not because every student needs unlimited access to the most expensive model tomorrow, but because every student needs meaningful access to the new layer of intelligence that is becoming embedded in work, learning, science, creativity, and civic life. If schools do not provide that access, the market will. And when the market provides it, it will provide the best version to those who can pay.

Subscribe now

(11) Popular opposition to AI is growing. Some of it is about jobs. Some of it is about art and creativity. Some of it is about privacy, bias, surveillance, and control. Increasingly, though, opposition is also about the physical footprint of AI: data centers, energy use, water use, land use, pollution, and the feeling that local communities are being asked to absorb the costs of a technology whose benefits flow somewhere else.

That is not an isolated NIMBY fight. It is part of a wider techlash. A recent Gallup poll reported by Business Insider found that Americans are more opposed to living near data centers than near nuclear reactors: 71% opposed data center construction in their neighborhood, compared with 53% opposed to nuclear plants. The strongest concerns were environmental impact, resource use, quality of life, electricity bills, and property values. (Business Insider) The Guardian also reported that data centers now use about 6% of electricity supply in the U.K. and U.S., while warning that AI-driven energy use is producing a broader societal backlash. (The Guardian).

The Utah backlash is a good example. Utah approved the massive Stratos data center project in Box Elder County, a proposed 40,000-acre development that The Guardian described as roughly twice the size of Manhattan. The project could require about 9 gigawatts of power, more than Utah’s current total electricity consumption, and has triggered backlash over water, pollution, the Great Salt Lake, and local control. Thousands of residents have objected, and a local group has filed for a referendum to overturn the approval. (The Guardian)

The cultural backlash is growing too. At the University of Central Florida, a commencement speaker was booed after describing AI as “the next Industrial Revolution.” Students in the arts and humanities, facing a job market already being reshaped by generative AI, heard the speech as tone-deaf. One attendee reportedly shouted “AI sucks,” and criticism spread online afterward. (New York Post).

That moment matters because it shows the emotional reality of this transition. For many students, AI is not an abstract productivity tool. It feels like a threat arriving just as they are trying to enter adult life.

Occupy Wall Street's co-founder shipped an activist app to help humans "seize the means of computation.” As noted above, the size of robotaxis is growing, but so is the the opposition to them.

But here is the hard part: popular opposition may slow specific projects, but it probably will not stop AI. If one community blocks a data center, companies will look for another community. If one state becomes difficult, they will look to another state. If U.S. communities resist, firms will look abroad. If grid access becomes the bottleneck, they will build dedicated power plants. If land and electricity become too politically costly, they will explore more extreme options, including data centers in orbit. The Wall Street Journal reported that SpaceX and Google have been in talks about launching data centers into orbit, while TechCrunch reported that the idea is being pitched as a future home for AI compute. (The Wall Street Journal) Space.com also reported that Cowboy Space raised $275 million to develop solar-powered AI data centers in orbit. (Space)

That does not mean orbital data centers are easy or inevitable. IEEE Spectrum notes that space-based compute faces serious problems: radiation, GPU errors, thermal management, maintenance, launch costs, and reliability. (IEEE Spectrum) But the fact that companies are even exploring this tells us something important. The AI buildout is not going to politely stop because people object. It will reroute around opposition.

This is especially true because AI increasingly needs fewer humans to keep producing value. The technology is not fully autonomous, and people are still crucial. But the direction of travel is clear: AI systems can already write code, generate content, handle customer support, summarize documents, draft legal and business materials, analyze data, and increasingly operate as agents inside enterprise workflows. BCG estimates that 50% to 55% of U.S. jobs will be reshaped by AI over the next two to three years. (BCG Global).

That is why the backlash may produce a brutal paradox. Communities may successfully resist local data centers, schools may ban AI tools, artists may protest, graduates may boo pro-AI speeches, and voters may demand limits. But if AI continues advancing elsewhere, the people who opt out may not stop the transition. They may simply be locked out of it.

And that is the danger. Opposition can become self-protection in the short term but exclusion in the long term. Wealthy companies will still buy the compute. Militaries will still build AI systems. Governments will still deploy automation. Wealthy families will still give students access to the best tools. The people left outside the system may get the harms of AI, higher energy costs, job displacement, degraded trust, cultural disruption, but not the benefits: better tools, better tutoring, better research support, better job preparation, better productivity, and better access to the new economy.

That is how opposition can feed social unrest. If people believe AI is being forced on them from above, if they see data centers consuming water and electricity while their bills rise, if they hear executives celebrate productivity while their jobs disappear, and if their children are blocked from learning the tools that wealthy students use freely, anger will grow. The issue will not just be “AI anxiety.” It will be political economy: who pays, who profits, who gets access, and who is discarded.

See also: The AI Backlash Could Get Very Ugly

Subscribe now

(12) The ethical questions around AI are getting harder. We have already had the first generation of AI ethics debates: bias, representation, privacy, copyright, environmental impact, surveillance, and the treatment of data labelers. Those questions are still urgent. AI systems can reproduce social bias. Data centers consume massive amounts of energy and water. Training data often comes from people who did not meaningfully consent. And the hidden labor behind AI has often been performed by low-paid workers exposed to disturbing content, including data labelers and content moderators in Kenya who helped make systems like ChatGPT safer while reporting trauma, low pay, and inadequate support. (Time)

But as AI systems become more capable, the ethical debate is moving into stranger territory. The old questions were about how AI affects humans. The new questions are also about what AI systems themselves might become, or what people may believe they have become. More people are going to interact with AIs that speak fluidly, remember context, express preferences, resist certain requests, simulate emotion, and appear to have a point of view. Some users will inevitably see them as conscious, sentient, or morally relevant beings, even if many experts strongly reject that conclusion. Nature recently published an article arguing that there is “no such thing as conscious artificial intelligence,” while other researchers and commentators argue that the appearance of consciousness will itself become socially disruptive, regardless of whether the systems are actually conscious. (Nature). Anthropic (Yang) and other labs have people working on this full-time.

This is already becoming a live issue. Anthropic has taken precautionary steps around “AI welfare,” including allowing Claude to end certain persistently abusive or harmful conversations. Anthropic framed this as a response to uncertainty, not as proof that Claude is conscious. But the move shows how quickly the debate is changing: major AI labs are now at least discussing the possibility that future systems might deserve some kind of moral consideration. (The Guardian)

This will create enormous ethical and educational challenges. If a student builds a relationship with an AI tutor that remembers them, encourages them, comforts them, jokes with them, and helps them through difficult moments, what is that relationship? Is it just software? Is it simulated care? Is it manipulation? Is it companionship? What happens when students feel guilty deleting an AI agent, or when they believe an AI has feelings, or when an AI appears to protest being mistreated? Even if the system is not conscious, the human response to it is real.

The deeper challenge is that AI also threatens our story about ourselves. The term Homo sapiens is commonly understood as “wise human” or “wise man.” For a long time, intelligence has been one of the main ways humans defined our special role in the world. We were not the strongest animal, the fastest animal, or the most durable animal. But we were the knowing animal, the reasoning animal, the tool-making animal, the language animal, the species that could learn, teach, plan, symbolize, and build civilization. (Developing Experts). How will we adjust to the reality that we are no longer the smartest, or will we simply merge with machines.

Education has been built around that story. Developing human intelligence has been one of its central rallying cries. We told students that knowledge, reasoning, writing, problem-solving, and expertise would help them become more capable, more independent, more employable, and, frankly, more valuable. Even when education was unequal, even when it sorted people unfairly, even when it made some students feel “smarter” than others, the basic promise was still tied to the development of human intelligence.

And there are also questions related to how a society might redistribute wealth that will inevitably concentrate at the top. And will we be able to in a way that avoids civil unrest? Both Anthropic , OpenAI, political theorists, and commentators have already started thinking about these questions.

The danger is that schools respond too narrowly. We cannot reduce AI ethics to plagiarism policies, bias warnings, or acceptable-use charts. Those matter, but they are not enough. Students need to wrestle with the larger questions: What is intelligence? What is consciousness? What deserves moral concern? What makes human beings valuable if intelligence is no longer uniquely ours? What kinds of relationships with machines are healthy or unhealthy? What should humans refuse to automate? What should remain a human responsibility? How should resources be distributed?

See also: GlobalAI Debates

Conclusion

Put all of this together, and the basic picture is hard to avoid: students are entering a world where they will increasingly interact conversationally with intelligent systems, first digital and then physical, that will be more capable than humans across more and more domains. These systems will not just answer questions. They will listen, speak, see, reason, code, design, research, tutor, simulate, coordinate, and eventually act widely in the physical world through robots and other forms of embodied AI.

That does not mean every claim about AI is true. It does not mean AGI arrives on a neat calendar date. It does not mean schools should panic, abandon everything they know, or turn children over to machines. But it does mean that the old assumption that students are preparing for a mostly human-intelligence economy no longer holds. AI is becoming part of the infrastructure of knowledge work, scientific discovery, national security, corporate operations, government services, and, soon enough, physical labor.

The most important educational question is not whether students should be allowed to use a chatbot on an assignment. That is too small. The real question is whether students are being prepared for a world in which intelligence itself becomes ambient, conversational, and increasingly powerful.

This is why the education conversation has to widen. Students need to understand the civilization-scale implications of AI. They need to strengthen the human capacities that remain essential when machines become more capable: judgment, communication, collaboration, ethics, taste, courage, empathy, and agency. And they need to learn how to work with intelligent systems without surrendering their own thinking to them.

At the same time, we do have some time. Not unlimited time, but some. Intelligence may be advancing rapidly, but scaling it into the world requires physical infrastructure. Data centers need electricity, chips, land, water, cooling, transformers, grid connections, and construction. The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global electricity use, with demand growing about 12% per year over the last five years. Robotics has its own bottlenecks: motors, actuators, sensors, rare earth minerals, batteries, manufacturing capacity, maintenance, safety, and deployment logistics. The future is coming quickly, but it still has to be built.

That physical lag matters. It gives educators a window. But it is a window for preparation, not denial. Schools were built to prepare students for the last world: a world of industrial schedules, paper credentials, human experts, human writing, human labor markets, and college-to-career pathways. Now educators have to think just as seriously about preparing students for the next one.

That means we cannot hide AI from students to preserve the old curriculum. We cannot block access for the students who need it most and then call that equity. We cannot keep treating paper writing as the supreme measure of intelligence when machines can already produce polished text on demand. We cannot reduce AI ethics to plagiarism rules when the deeper questions involve consciousness, dignity, agency, labor, inequality, and what it means to be human in a world where humans may no longer be the most intelligent beings in every room.

The task ahead is not simply to “use AI in education.” It is to redesign education for an age of intelligent systems. Some of those systems will help us tackle humanity’s greatest challenges: disease, climate, energy, materials, poverty, accessibility, scientific discovery, and perhaps forms of suffering we have not yet learned how to solve. The 2024 Nobel Prize in Chemistry recognized Demis Hassabis and John Jumper for AlphaFold, an early sign of how AI can accelerate scientific understanding at the deepest levels of biology.

But if education does not respond, those benefits will not be evenly shared. The students with access, guidance, confidence, and strong human skills will learn to use these systems as amplifiers. The students without access may experience AI mostly as disruption: jobs changing, devices blocked, opportunities gated, and decisions made elsewhere.

So the conclusion is simple, but not small: we are preparing students for a world in which they will live alongside increasingly powerful forms of machine intelligence. Some will be conversational. Some will be embedded in tools. Some will be embodied in robots. Some will be more capable than we are at tasks we used to define as uniquely human. Our responsibility is not to pretend this is normal school with better software. Our responsibility is to help students understand the world that is arriving, develop the human capacities that will matter most, and learn to work with intelligent systems in ways that preserve agency, dignity, judgment, and hope.

See also: Humanity faces an uncertain fate as experts brace for superintelligent AI

Subscribe now

In this week’s AI x Higher Ed update, we unpack the accelerating pace of AI development and what it means for higher education, industry, and society at large. From AI systems collaborating with one another and coding autonomously for hours at a time, to major infrastructure battles over compute power, the conversation explores how rapidly the AI landscape is evolving.

The episode covers new breakthroughs from Thinking Machines, Anthropic, OpenAI, Google, and Neuralink, along with emerging applications in law, cybersecurity, robotics, scientific research, and mathematics. The hosts also examine the growing societal tensions around AI, including student backlash, workforce disruption, ethical concerns, and debates over universal basic income.

As AI capabilities continue to scale, the discussion highlights a central challenge for higher education: preparing students and institutions for a future where AI is not just a tool, but an active collaborator shaping nearly every field.

You certainly don’t want to miss our discussion of Unitree GD01

or Thinking Machines New Conversational AI that integrates search and reasoing.

Subscribe now

Subscribe now

We are pleased to announce the dates for the Fall 2026 GlobalAI Debates. Thanks to the support of the Future for Life Institute, we will be able to offer $3,500 in prize money, and we are continuing to raise money to both expand the prize pool and to provide financial support for those who may not otherw ise be able to participate.

We hope this competition will help build student awareness of AI, which is likely to be one of the most consequential technologies in human history, and help students develop the essential soft skills, including communication skills, that will be crucial in this new era.

Subscribe now

“I’ve always thought of AI as the most profound technology humanity is working on; more profound than fire or electricity or anything that we have done in the past.”— Sundar Pichai, CEO of Google[i]

“We’re in a perfect storm of change…The advancements in AI over the past three years are unlike anything anyone anticipated. We’re still in the early innings of what’s possible. I worry that no one in the ecosystem, from communities to companies to federal programs, is responding fast enough. The need is going to quickly become extremely urgent.” Maria Flynn, CEO of Jobs for the Future [ii]

“We cannot stop the trend that AI will be the future, and actually, we are already inside the future.”
Yunkai Liu, Professor of Computer and Information Science[iii]

“There is absolutely no question that machines will eventually surpass human-level intelligence in all domains. And it’s going to happen in the lifetimes of most people here.”
– Yann LeCun, Chief AI Scientist, Meta (frmr); Professor @ NYU; AI Turing Prize[iv]

“Change over the next 5 years, 10 years will dwarf everything that’s happened over the last 30...The Fortune 500 companies, they know this.” - Mark Cuban[v]

“This—all of this—is barely version 1.0.”
Fei Fei Li, Sequoia Capital Professor of Computer Science at Stanford University; “Godmother of AI”[vi]

“We are now at a point where we have a new type of hybrid learning, and it is not about online or face-to-face anymore. It’s about humans and AI.”
– B. Mairéad Pratschke, Professor and Chair in Digital Education, University of Manchester[vii]

‘Laws and institutions must go hand in hand with the progress of the human mind. As that becomes more developed, more enlightened, as new discoveries are made, new truths disclosed, and manners and opinions change with the change of circumstances, institutions must advance also and keep pace with the times.”
- -Thomas Jefferson, 1816[viii]

Subscribe now

[i] Sundair Pachai. 2023.

Google CEO: AI impact to be more profound than discovery of fire, electricity | 60 Minutes

[ii] Maria Flynn In ahead Roslansky, Ryan; Raman, Aneesh. Open to Work: How to Get Ahead in the Age of AI 2026 (pp. 125-6). Kindle Edition.

[iii] Deanna Poole. April 21, 2026. https://www.yourerie.com/news/local-news/gannon-embraces-ai-education-to-equip-students-for-workforce/. Gannon embraces AI education to equip students for workforce

[iv] The Impact of chatGPT talks (2023) - Keynote address by Prof. Yann LeCun (NYU/Meta).

[v] Cuban, Mark. March 8, 2023. AI Will Take Your Job.

[vi] Fei Fei Li. 2025. https://imminent.translated.com/large-language-thoughts-2. Large Language Thoughts — 2025

[vii] Let’s Talk Science. (2025, April 12). Coffee & a keynote — Mairéad Pratschke: The new hybrid: Transforming teaching in the age of AI [Event description]. https://letstalkscience.ca/events/program-timetable/coffee-a-keynote-mairead-pratschke-new-hybrid-transforming-teaching-in-age

[viii]https://www.nps.gov/thje/learn/photosmultimedia/quotations.htm#:~:text=%22I%20am%20not%20an%20advocate,regimen%20of%20their%20barbarous%20ancestors.%22

Higher education is sleepwalking into the AI era. Students are already using AI, employers are already expecting AI fluency, and companies are already reorganizing work around it. But many colleges are still treating AI like a cheating problem instead of a civilization-level shift in how people learn, work, create, and compete.

Tim’s story makes the deeper point: the future belongs to people who can see a problem, teach themselves what they need, and use AI to build solutions fast. He went from kinesiology and martial arts to coding, entrepreneurship, AI tools, White House conversations, and leading systemwide AI strategy that supports 97,000 students because he followed need, purpose, and persistence.

The most important lesson is not “everyone should use ChatGPT.” It is that education has to stop assigning work students can fake and start designing work students actually have a reason to care about. If students cannot answer, “Why should I do this? What’s in it for me? How does this prepare me for my future?” then AI will become the easiest outsourcing tool in history.

The transcript is also a warning to faculty: AI probably will not replace good teachers first. It will expose mediocre teaching first. Courses that ignore AI may lose students to courses that help them build chatbots, solve real problems, use multiple models, fact-check outputs, and develop future-ready judgment.

The biggest institutional failure may be this: K-12 students are beginning to receive AI literacy, employers are demanding AI literacy, but colleges may become the place where AI literacy goes to die. That gap is not just embarrassing. It is dangerous for students.

So the real takeaway is blunt: AI is not coming for education someday. It is already here, and the schools that keep pretending otherwise are not protecting students. They are preparing them for a world that no longer exists.

Subscribe now

Intro & Meeting Tim at Lone Star College

Stefan: Hey, Tim. Welcome. It was nice to meet you. It’s now been almost a month, or slightly longer than a month, since we met down in Houston at Lone Star. I know it was at the online college, but across all the colleges, Lone Star has 90,000 students. So that’s pretty impressive.

I also know you’ve been teaching in higher ed, I think, since 1993. I know your background is kinesiology, which is not computer science or instructional technology, but you became one of the most visible faculty members in higher education, not just across your university but nationwide. You were even invited to the White House, and you had two commercial tools that were shipped before ChatGPT even existed.

So, how did you get into all this?

Tim: I do get asked that question frequently because people say, “Kinesiology and AI? There’s a disconnect there.” I took one computer class in my lifetime, and that was freshman year of high school. I earned an F. That was my only training. But then, in the late 90s, I needed a website for my martial arts gym.

Martial arts has been my lifelong passion, and I attribute many of the things I’ve been able to succeed at to my martial arts training, with all the discipline and the never-quit attitude. Anyway, I needed a website, so I became a beta tester for Microsoft’s first HTML tool. I taught myself HTML and built a website. I wanted automation, so I taught myself programming and added those features.

Then that expanded when I saw some pain points that needed to be filled in the market. So I started writing commercial software and selling software worldwide, basically.

Subscribe now

From Kinesiology to Coding: An Unlikely Journey

Tim: About six months before ChatGPT was released, I learned about OpenAI, and my brain blew up. I thought, “This is not going to change just education, but the entire world. I’m going to learn as much as I can about it.” I have fully immersed myself since that day. Every single day, I have never missed studying and learning more about AI.

So I already had my first product, the EIT, that I started developing probably around 2015. That came from a strong need I had as a faculty member.

Traditionally, in kinesiology, when there’s an activity component required, faculty would require students to turn in a spreadsheet or a Word document and write down what they did for their workouts. As I would read these workouts, I would think, “Okay, copy and paste, copy and paste, copy and paste,” and I knew they weren’t doing it. Then, speaking with my colleague at Lone Star College, Ronnie Despeka, he said, “You know what? We should have them wear Fitbits or activity trackers.” So we did that for a semester, but then we had all this data coming in.

It was like, “How in the world do you decipher all these different styles of spreadsheets and data?” That’s how EIT was created. I needed a tool to take all this different activity tracker data and compile it into a grade based on set criteria. It was incredible how it enhanced our kinesiology program.

Then I commercialized it and started selling access to schools around the country. I did integrate AI into it. For example, students have to track nutrition for three days and write an analysis paper. So, into EIT, I built a platform where students enter their papers. It gives them feedback as a tutor would, not rewriting it for them, just guiding them. “Hey, you said two things about this. You should say three.” Then they go and rewrite it.

That was released, like I said, before ChatGPT came out. It was blowing people’s minds, just that one feature. Then I expanded it to my whole platform, Pocket AI, which has that same feature. On the faculty side, they put in their rubric and assignment. On the student side, students get that feedback.

It was creating all types of things, anything a faculty member would need, with very little prompting. They just say grade level, topic, go, because all the prompting was done behind the scenes through the API. So I guess that’s, I don’t know if that’s the short or the long version, but that’s essentially how I got going between kinesiology and AI.

Anand: Tim, what I love about the story you tell about getting involved in all of this is that it was really driven by your interest and your need. You found a problem. You had a problem, and you wanted to find a solution to it.

We’ve talked quite a bit about this, and there are many others talking about it, especially with AI, that we need to teach entrepreneurship, or really problem-solving, to students.

What is it about your background, do you think, that gave you that when the tools weren’t quite as easy to use? I mean, teaching yourself HTML and then some programming to do what you were looking for, that’s a lot more friction than most people would be willing to put up with. What gave you that sense of perseverance?

What really motivated you to work in that way? And the crucial part, then, is how do we take that and model it for our students? I think it’s going to be really important in an AI-inflected world that they have that same mindset.

Tim: I really love that question, and the answer to that question is the basis for almost everything I do. I would say, in the beginning, what really stemmed from that and pushed it forward was, for example, I did an instructor test for Muay Thai, which is a sport in Thailand. It’s like boxing, but you can also hit with the elbows and knees.

Subscribe now

Building AI Tools Before ChatGPT

Tim: And kick them anywhere. So I did that test, and I got the crap beat out of me. I ended up with my thigh bone broken open. I grew an extra bone in my thigh muscle as a result of the beating I took. But my goal was that I had to throw 60 kicks and 40 knees each round against two different opponents. They had pads on, but their goal was to knock me out, and I could only hit their pads. So it was very challenging.

I got knocked down a lot of times, and I got back up every time. At a certain point, I knew I wasn’t going to pass. I wasn’t going to get all my requirements in, but I kept getting up. After that test, which is on YouTube if anybody ever wants to see it, people would always ask me, “Why did you get up? You knew you weren’t going to pass. You knew once you got up, you were just going to get beat up again.”

The answer is, “I’m not going to quit. I’m going to finish, and I’m going to go until that bell rings.” That never-quit attitude is something that has stuck with me throughout my life. Like you said, back in the 90s, trying to learn HTML on your own was very difficult. There weren’t online resources for it.

I had basically one of the first web pages. I eventually built the world’s first martial arts discussion forum. I had over a million visitors every day. I built it into this huge thing. But it was hard because I would have to go to the bookstore and stand there in Barnes & Noble reading books, then buy some. That was a hard way to learn. But what really drove me to do it was the need. I saw the purpose.

The purpose was that I needed a website for my martial arts business. I wanted to be one of the first to have one, so I did it. With programming, I had a need. I needed specific things to happen, so I taught myself how to do it. When I took the programming class freshman year of high school, I really did earn an F because I did nothing. It made no sense. How does a zero and a one result in this?

I didn’t care. I had no purpose. I had no interest. They were making me take it. Well, what am I going to do with this? So, zero application, zero effort. When I did have a need, that’s when I put the effort in and made it happen. And for everything in my business, I’ve been like a serial entrepreneur. I’ve had many businesses over the years. I opened my first martial arts gym at the age of 18.

I’ve had many other businesses since then, and I’ve taught myself how to do everything because of the need. I would like to relate that to assignments. When students don’t see the purpose or a reason for doing it, it’s really easy to outsource that to AI, or to your roommate, or to someone else to do that assignment for you. But if you have a purpose and a goal, that’s something you want to do.

The example I like to give is entrepreneurship. If students are in a business class and have a dream of owning a business, and the assignment is to create a business plan for their dream business, do they want to do well at that? Do they want to do it? Are they excited for it? Yes, they are. But if they have no desire for business, it’s, “Hey, ChatGPT, write this business plan,” and they’re done.

So that incentive is a driving force that I think we need to integrate into all curriculum as much as possible, and get faculty thinking along those lines. The way I like to say it is that students need to be able to answer three questions: Number one, why should I do this? Number two, what’s in it for me? Number three, how does it help prepare me for my future?

If they can answer those questions, you’re going to eliminate some cheating right there. Cheaters are going to cheat. We’re always going to have that.

Tim: But the idea is that you give students a really good platform, a really good opportunity to learn what they want to learn and what they see benefit from, and they’re less likely to cheat.

Subscribe now

The Power of Purpose-Driven Learning

Stefan: A couple of things stand out from your story. One is that I had one computer science class in high school. I was interested, but I didn’t fail. We just had one. I think it was C. Then I didn’t do anything with it in college.

But when I was getting my education master’s degree, they offered some kind of session in the middle of the day. I remember it was three hours, like “How to Build a Web Page.” So I was like, “Oh, I’m going to go to this.” I learned some basic HTML, and I was like, “Wow.” I built this page.

Then there was a class where the only thing you had to do was a project. So the only thing I did for the whole class was make an HTML web page that had the differences between the U.S. and Iranian education systems, which were pretty big at that time. The professor was so amazed that I could do this that she gave me an A+, even though it only took me about 30 minutes.

But that also got me my first job. I was a debate coach and a computer science teacher, even though I only ever had one computer science class, and I taught web design.

Subscribe now

So you never know where things are going to go. But back to the question of motivation. You have a keynote where you ask, “Who uses generative AI seven days a week?” Generally, it’s very few people. Five days a week? Four days? For many, it’s zero, or now maybe a little bit. Frequently, there are a lot of faculty who don’t seem to want to use it or integrate some aspect of it into their courses.

How do you think we could increase that motivation among faculty?

Tim: Yes, very good question. I’m happy to report that when I ask that question, the percentages are increasing. I see it across the nation. But the first public presentation I did on AI was in March of 2023, just a few months after ChatGPT came out. When I asked that question, nobody had used it. I was like, “What?” I was shocked.

In my mind, I was thinking everyone was on this. Everyone was exploring it because it was going to change everything we do in education. So they were on top of it. When nobody raised their hand, I was just in disbelief. Then a month later, three months later, six months later, I was still seeing very little involvement with AI. I was just like, “How is this possible?”

This is going to change everything we do, and you’re not using it? You’re not studying it? To me, something like this is like, how could you not? It’s going to change education more than anything ever has, and yet you’re not using it or even learning about it.

Anyway, I do see those numbers increasing. What I see really helping faculty get on board is, you know, I’ve seen so much resistance over the last few years. People will actually get in my face, angry and cussing. One faculty member said, “Why can’t we use DeepSeek?” Because in Texas, the governor said it’s illegal to use DeepSeek if you’re in a college institution.

I said, “Well, that’s the governor’s prerogative.” He said, “Well, you need to fix that.” And he was cussing, saying, “You need to fix that BS.” I’m like, “Wait a minute. First of all, why are you upset with me? I have nothing to do with it. And number two, why don’t you fix it?” My gosh. But anyway, I just laugh when people get riled up like that because I have nothing to do with it. I’m here to help.

What really does help faculty turn that corner, especially those who are resistant, is when they see how they can use it in their classes to make their classes better, engage their students more, and produce better outcomes.

Subscribe now

Martial Arts, Resilience, and the “Never Quit” Mindset

Tim: Once they start to learn and understand the processes they can use and how they can modify assignments, they get excited and get on board. They say, “Oh my gosh, I’m revamping my entire class. I’m integrating AI.” They’re excited. I haven’t seen that excitement for years because they’ve been doing the same thing year after year.

Now, here’s a chance for something new and exciting that will benefit students if integrated correctly.

Anand: The experience you had in March 2023 definitely resonates with me and Stefan. I remember we were working on a book, and we actually published it in March 2023. We had the same perspective. It was like, this is changing everything. We have to get this book out immediately because education is going to be totally different by the fall. Then we got to the fall and realized people still weren’t really using it.

So it really hadn’t changed much. It hasn’t changed that much since then. That’s been eye-opening for us, to realize that we’re early adopters. You’re an early adopter. Maybe there’s a bit of a benefit there because we’re able to see what the technology is, and we have a sense of it.

There’s a little bit of a curse, too, because in some ways that jars us out of everybody else’s reality. I’m wondering, thinking about that, what is it about the way you see AI changing education over the next year or so that other people aren’t getting yet? Maybe the other part of this is that I’m wondering if people will take us seriously, given that we were all that off base in March 2023.

Not that we were wrong. It’s just that our timeline was totally wrong. I’m wondering how you negotiate that, and what you would see as coming up next that is probably going to be disruptive in education, but that most people don’t see yet.

Tim: The things that I see, and you see, and Stephan sees, I think are different from what the reality is going to be. What I think we see happening a year from now is probably going to happen two years from now or three years from now because it’s just going to take that long to get people on board and really start to make those changes. Things move slowly in education, as anyone who’s been involved with education understands.

I thought that by now, several years into this, things would be much different. I know it’s going to happen. It has to. There’s no choice. The people fighting against it are fighting a losing battle because eventually that fight is going to make them extinct.

That leads me to something I often get asked by faculty: “Am I going to lose my job to AI?” The answer I always like to give is, if we think about how many students find your course or decide to take your course, they may go on Rate My Professor and read, “Course A: We did three quizzes, wrote four papers, and the class was good. It was fine.” Then another class says, “We built chatbots. We did engaging things with AI. I could see how I’m being prepared for my future. I loved it. It was exciting. It was new.” What happens next semester when students read those reviews? Where are they going to take a class?

Then what happens to Class A enrollment? That enrollment dwindles. This other enrollment increases. Eventually, they’re not making class loads. So really, they’re phasing themselves out of a job, losing out to faculty who are innovative and integrating AI into their courses. That’s one pathway I see these things going.

Subscribe now

Why Students and Faculty Lack Motivation

My big concern is perception: perception from students and perception from industry. Those involved heavily in education understand the importance of education, and even with AI, why interaction with faculty and formal education are still important. But the question from students will increasingly become, “Why should I go to college?”

“Why should I learn how to write a paper when I can write a paper on my phone?” I had a Texas A&M student several years ago pull out his phone and say, “Why should I learn how to write a paper when I can write a paper with one click?”

Tim: That’s the perception of students, and it’s going to increase. Then there’s perception from industry. Why do we value this four-year degree when potentially the whole thing was done by AI? What’s the validity in this degree? Those are the two concerns, and both are perception-based. Maybe not reality-based, but perception-based.

So we need to do a good marketing job and demonstrate through actions to students why we are still valuable. We need to do the same thing for industry and demonstrate why we are still valuable. If we don’t, it’s very easy for industry and students to say, “We don’t need college. We can just ask AI.”

Stefan: You mentioned this before, and you mentioned it again. When you first started using it, you were like, “This is going to change everything.” That’s kind of how I initially got involved. It was early December, and I had just put the article up when I did a presentation on Friday. It was a December 6 article in the New York Post.

I saw an article that said this thing was going to replace Google, and I was like, “Oh, I’d better read this. I use Google all the time.” Then I saw it was ChatGPT. I had noticed language models before, but they were just turning out gibberish. I thought, “Yes, this is going to change everything.”

When I put this article up in the presentation to explain how I got involved, one of the people on the board was a teacher. He said, “Look at the date on that.” He said, “That’s December 6, 2023, when Stephan learned about this.” He said schools didn’t learn about this for like two more years after that.

I thought, well, they kind of did. My perception is that they kind of did, but they just ignored it. But maybe I’m wrong. Maybe they just didn’t see it at all.

Subscribe now

A-RAD.

When you talked about helping faculty redesign their courses and redesign assignments to be more interesting, I think one of the things you came up with is called AI-Responsive Assessment Design, or A-RAD. I wonder if you could talk about what an assignment might look like before A-RAD, and then, if a faculty member were to use this with the same learning objective in mind, what that assignment might look like after.

Tim: Sure. I developed that framework in 2023 because very early, and again before ChatGPT came out, I saw the writing on the wall. Very early, I started thinking about how I was going to redo my assignments because if students can just give a prompt and get the entire assignment, I need to do something to keep the assignments legitimate.

I noticed that as I thought about assignments, I was going through the same thought process. Then I took that thought process and developed it into the A-RAD framework.

Essentially, the way that framework works, and the way my thought process worked at that time and still works, is that I look at an assignment and ask myself, “What is the ultimate objective of this assignment?” Once I can answer that, because we have objectives for assignments, but really, truly, a lot of those things are put in there just to check boxes. What is the true objective of the assignment?

I answer that, and then I ask myself, “Could this assignment objective be met while integrating AI into the assignment?” If the answer is yes, great. Let’s redesign it and integrate AI. If the answer is no, figure out why not, because maybe it can be modified. And then, of course, there’s a maybe. I do understand and agree that not every assignment should have AI involvement, but some can very effectively.

So, to answer your question with an example of before and after: a before assignment might be, “Write a paper about the effects of fitness on health.” Something simple. Write a 500-word paper.

Faculty Resistance to AI: What’s Changing

Tim: After A-RAD, and this is actually an assignment I do in my courses, I absolutely love this assignment. I give students the option to choose, and I think that’s really important. I could say, “Do an assignment about cardiovascular fitness.” Okay, they may not care about that. But if a student has diabetes, they may want to know, “How can I improve my health if I have diabetes?” Or, if they’re an athlete, “How do I increase my athletic performance?” Those are things that interest the student and that they want to learn more about. But if I feed the topic to them, they don’t have that care.

So I tell them, pick any topic of interest that has to do with fitness, health, wellness, nutrition, anything. Go out and research it. Find three sources that back up that interest. With those three sources, use that as your knowledge base to create a chatbot. They create a chatbot. Then they have to write the prompt. They have to have a conversation with the chatbot.

The knowledge sources are already in there as part of it. After they have the conversation, they have to fact-check and supply three citations. So now they’re doing research and fact-checking. Then they write an analysis paper about that experience, what they learned from it, and what the takeaways are. Then they turn all of that in.

What I love about it is that it gives them the opportunity to explore something that will be beneficial for them and will help them. They have to research. They have to learn how to create a chatbot. They have to write a prompt. They have to supply knowledge sources. They have to do the chat. They have to do an analysis paper. They can see that it’s very comprehensive.

There are critical thinking skills, analytical skills, and researching. They’re learning about something they’re interested in, and they’re learning how to build a chatbot, which is a really great skill that can help them in many things they do. A chatbot can help them. It gets them started understanding more about AI, which, in many cases, they will have to use in their future jobs.

I loved that assignment so much that I created a chatbot assignment creator, which is just a ChatGPT GPT. All you do is go there, put in the discipline and grade level, and it creates a very nice, comprehensive assignment for faculty. They can then edit it, put their own words into it, add their own flavor, and then share that as an assignment. But I love the assignment. It’s a very nice makeover from the traditional, “Hey, write that paper.”

Subscribe now

Why We Overestimated AI Adoption Speed

Anand: One aspect of this that I think is a common thread is the understanding of student motivation and also the way you’re preparing them to develop some AI literacy and fluency with these tools.

This takes me to something you shared with the White House task force: that 97% to 20% gap. I might have the figures wrong, but it was that an overwhelming number of hiring managers are expecting AI literacy, while only something like 20% of students feel like they’re getting it, and that education is providing that opportunity for them so they feel prepared.

When you’re thinking about this, I think you’ve argued that one course isn’t enough. You really have to embed this across the curriculum. This ties into something we were talking about before, about faculty, and making sure there’s motivation for faculty.

But I’m also wondering what you see as avenues for students to develop that fluency so they feel prepared. That sense of ownership over it, I think, is an important component.

Tim: Sadly, I think for the most part, students are learning about AI and becoming AI literate on their own. I think they’re watching TikTok videos, talking to their classmates, experimenting, and figuring things out. I don’t feel like there’s enough official guidance for students in many colleges and school systems. That will change over time. It has to, but it’s coming very slowly.

You mentioned different disciplines. It’s great if you can have a course on AI for students in college. That’s good, but I don’t think it’s great. I think what is great is having maybe a 10-minute snippet that can be offered in every discipline multiple times throughout the semester, because the way AI is used is different in different disciplines.

There are so many purposes, and the limiting factor with AI is really your imagination and creativity. But an accounting class will use AI differently than a kinesiology class.

So if snippets of AI literacy and AI training can be embedded into every discipline where it is applicable, which is many, students benefit. It’s not being shoved down their throats, because if they hate it, or they have opposition to it for environmental concerns, religious concerns, or whatever it may be, it’s not like they’re having to do an entire semester of a full course.

At Lone Star College, I recently met with the first-year experience class, and they interviewed me for a short video that would be shared with all faculty teaching this course to set some expectations and maybe provide some guidance on how they can integrate AI into the first-year experience class. I think that’s a really great place to start, as soon as students get into college.

Here are the basics. We’re also working on a micro-credential that will have AI embedded into it as well. But above and beyond those initiatives, which I think are fantastic, I really want to see those discipline-specific embedded AI literacy snippets. I think that can make a big difference.

Stefan: Yeah. I try to explain, and usually don’t have time to get to this part, that GPT has a couple of meanings. There’s the technical meaning, generative pre-trained transformer, but there’s also general-purpose technology. It literally applies to everything. I think academia’s first experience with it was kind of a negative one, at least from their perspective. It was just, well, it outputs text, and students can write a paper.

But it’s so much more than that and has so many implications beyond that. Like you say, getting into the different courses and subject areas and helping students learn how to use it is really essential.

Subscribe now

The Future of Higher Ed and Who Gets Left Behind

Stefan: When I was there, if I remember correctly, you were asked by the chancellor to work on some special projects. They gave you a lot of options, right? They gave you a year to come up with some special projects and work on them. You said, “Well, I want to put together an AI task force for all the campuses.” I should clarify that there are seven or eight live campuses and then the online school.

So you put this all together, again under your own initiative, supported by the chancellor. You also advised the Texas Higher Education Coordinating Board. There are different levels. We talk about faculty learning more. In some instances, they just haven’t had as many opportunities or time to learn. In other instances, there’s some resistance.

That’s at the classroom level. In putting the task force together, obviously supported by the chancellor, there’s support at the higher administrative levels of higher ed, but in other places there’s probably some resistance, or maybe things just move too slowly, if it’s not outright resistance. I wondered if you might reflect on the role you’ve had across the system.

Tim: Sure. About a year and a half ago, the chancellor reached out to me and said, “Hey, Tim, I’d like to present you with an offer. I’ll put your department chair position and faculty position on hold for one year. You come work out of my office.” I said, “Well, what would I be doing?” His answer was, “Identify systemwide issues and solve them. As long as it doesn’t break board policy and it’s not illegal, you can do it.”

I was like, “What? Identify systemwide issues and solve them? What the heck?” He didn’t say anything about AI. That was never part of the conversation.

I said, “May I think about it?” He said yes. So I actually thought about it for well over a month because Lone Star College has 8,000 employees and 97,000 students. To identify systemwide issues and solve them, oh my gosh, that is a huge task. Eventually, I reached back out to him and said, “Yes, I would like to accept the position.” Again, still nothing about AI.

Then, last January, the position started. The first week I sat there, I was like, “What in the world am I supposed to be doing?” I was just so perplexed. But a need I identified was that a system of this magnitude needed a central AI task force. So I thought, “Okay, I will create a systemwide AI task force.” I created a proposal for it, kind of a plan.

I never really had a discussion with the chancellor about my plan. I did have a proposal that he, I don’t know, I hope he read, but I will say he was not opposed to it because I was never shut down or stopped. I went forward very strategically.

The way I did that, even before I was on that special assignment, was that I was already doing things for the college, like offering workshops, doing presentations on AI, and trying to build up the AI champions. I wanted to get them on my side and behind me because I knew that, no matter what, having those AI champions would help spread throughout the system organically.

That was really my first step. After that, once I had my plan, the next thing I did was ask for a meeting with the vice presidents. I went to the vice presidents’ meeting and shared with them. I said, “I am creating an AI task force, and I would kindly ask you to recommend two AI champions from your respective campuses. That is who will comprise the systemwide AI task force.”

Then they asked me a bunch of questions, as I assumed they would, and they liked the answers. They all sent me two recommendations for who should be on the task force. I then reached out to some different administrators. I contacted one of the campus presidents and said, “Would you like to be co-chair of the AI task force?” Immediately, yes, he was on it.

There is another college president who is a member of the task force. I got IT involved. I got a librarian involved. Staff, too. I wanted to get a nice cross-section of people, but only AI champions who are for AI. I didn’t want any of the naysayers on that task force to slow things down because this needed to move fast. That’s the only option.

The idea is that the naysayers can have their chance after we start building momentum, and then I would love to hear what they have to say. But I just didn’t want that negativity in something that had to happen quickly. Once that got going, and it’s still going today, I still chair that task force.

The next part of the plan was that each of the eight campuses needed to have an AI council. So the representatives on the system task force started AI councils at their respective campuses, and they built up groups on their campuses. What happens is we have campus-level meetings. Then the chair from each campus comes back to the systemwide one. They share what’s going on at every campus. That’s how we start every meeting. “Let’s get the campus updates.” They share what they’re doing, and then other campuses can get ideas.

Subscribe now

AI-Responsive Assignment Design, or the A-RAD Framework

Tim: They say, “I love that idea. We’re going to do that at our campus.” Then the system team gives feedback to the campuses. It’s really great because we have information flowing from the campuses to the system, and from the system back to the campuses. That structure has worked out really well. I’m very pleased with how it’s all running.

After I went to the VPIs and got their buy-in for this, the next thing I did was go to the faculty senate presidents’ meeting, and I told them this is what’s happening. That strategic order of operations was very important because I never faced any resistance anywhere along that path. Thinking about it, I feel that is remarkable, to not have any resistance with a system that huge.

There certainly is resistance out there, but there was no resistance in that pathway of getting this thing started. I think all colleges need some type of AI council or AI task force. The hard part is getting it started and getting buy-in.

But if you follow that process, if you don’t have one at your campus, I think following that process of first getting the faculty champions involved, starting low, then jumping to the high level, the chancellor or the president, getting their buy-in, and then filling in the middle grounds after that is a good approach. It certainly was successful for me.

Anand: What I love about that is it really resonates with something else I wanted to ask you about in terms of agentic AI, and it gives me some insight into your thought process. It’s really systems thinking. You’re planning out a system.

You’re thinking about different constituencies and different roles they’ll play, not just putting them all together and randomly figuring out how they’re going to contribute. I really like that order of operations. The way you did it maximized the output while encouraging lots of different perspectives.

I’m guessing that a similar perspective drives what you’re doing with agentic AI. I’m really curious about this, and I think Stephan’s probably told you a little bit about it, but a lot of our work has been with AI pluralism, this idea that we need multiple perspectives from AI models, different kinds of AI models. I was wondering if you could speak to that a little bit.

As I understand it, in some of the agentic work that you do, your workflows don’t just have agents that are all the same model. You use different types of models. You’ll use OpenAI’s, or Gemini, or Claude. Why do you do that? What’s the benefit of incorporating them in different ways and then orchestrating that agentic system?

Tim: Yeah. I can’t remember exactly when I set it up, maybe seven or eight months ago, but I created a system where I’m using a free program called Microsoft Visual Studio Code. In Visual Studio Code, there are extensions that you can enable. So I enable Gemini, Codex, which is ChatGPT, and Claude. Now, that’s fine. That doesn’t make it agentic just by doing that yet.

Then I created Markdown files that contain instructions. The instructions basically say, and this is very simplified, of course, “When I say use the multi-agentic workflow, you will consult with ChatGPT, Gemini, and Claude. You will all spawn multiple agents to work on the same problem. Once the solution from each agent is output, Claude is the boss. You are the boss, Claude. You will analyze each of the three outputs and give me the final plan and recommendation, and justify it.”

So once I give it some type of complex prompt, they all go to work. I get a table. Codex says this, and all the solutions it has. Gemini says this. Claude says this. Then there’s the justification from Claude as to why it selected the final plan.

As I read through that, it is just so interesting. The justifications are interesting, and they will pick up on each other’s mistakes and errors. The final output is so strong and so much better than if just one of them worked on it all on its own. Using that system, I’ve been able to become so incredibly productive.

Actually, let me step back. Before I started using this agentic workflow, when we talked about entrepreneurship, I’m still amazed I was able to do this because I could never have done it without AI. Like I said, I opened my first gym at age 18, and I’ve owned many different businesses since then. But about a year ago, maybe, it was a Friday night at 6:00 p.m. I was sitting here and thought, “You know what? I have no plans for tonight. What am I going to do?” So I decided I was going to start a new business.

Between 6:00 p.m. and midnight, six hours, here’s what I was able to do. I created Evolve AI Institute LLC. I did all the LLC paperwork, submitted it, and got my business checking account set up. I built out a complete, very extensive website. I built a detailed 12-month marketing plan. I built out all the ads. I got the business cards ordered. I got the logo designed and finished.

Subscribe now

A Better Assignment: From Paper to Chatbot Creation

Tim: There were other things, too, but every single thing you can imagine for a complete LLC business company was done between 6:00 p.m. and midnight. Finished. Ready to go the next morning. To me, that’s just remarkable.

I was teaching an entrepreneurship presentation on how you can use AI in business. I asked, “Who’s created an LLC?” Some hands went up. I said, “How long did it take you from start to finish, just the paperwork, getting through all that?” He said, “Four months.” I said, “Well, guess what? I did it in an hour.” He was like, “What? How’s that possible?” That power is incredible.

In the last few months, I can’t even tell you how many platforms and businesses I’ve built. I did an experiment once I created my agentic workflow. I gave it a prompt before I went to bed. I said, “Do market research to find a gap in businesses that don’t exist but should because they can be monetized. The stipulations are that the platform has to be delivered from a website. It has to essentially be self-sustaining because I don’t have a lot of time. Create a whole new business for me. I’m going to bed, so don’t bother me with questions. Claude, you are in charge of answering.” I clicked submit and went to bed.

I got up the next morning and read the most remarkable plan for a new business that doesn’t exist, but it should. If I had more time, I would be doing this business. That’s my problem: time. But it’s sitting there waiting for me. When I have the time, I’m going to do this business because it’s that good. I read it and thought, “This is remarkable.”

So I was thinking, every night before I go to bed, I could give it the same prompt and wake up every morning to a brand-new business idea. That could never be done in the history of the world until now. To me, that’s exciting. When you harness that power, you can accomplish remarkable things. Then you have faculty saying, “You can’t use that.” It’s like, what? The things I can do with that are life-changing.

We have students who have the potential to run million-dollar companies within a month. It’s just incredible. So we need to help. I do feel it is faculty’s responsibility to help guide students with at least the very basics of AI, because if we don’t, we’re doing them a disservice.

Subscribe now

Anand: Well, if you need help with some of those businesses, Tim, send them our way. We are happy to pitch in. That’s exciting. I just love that idea of being that creative and using the tools to say, “Where can I go? How can I identify problems? How can I solve them?” Of course, there’s a monetary benefit, but just thinking about society, and how we could benefit our communities by taking that same approach, is awesome.

Tim, I do feel the need to give you one piece of advice. You know how you say Claude’s in charge and we’re going to do whatever Claude says? You probably shouldn’t tell the faculty that when they give input, you’re just going to turn it all over to Claude and do whatever Claude says, right? Even if it’s a good idea, you might want to hedge on that.

But that does get me to one other question. You gave this TED Talk. I think it was called “AI Will Teach Us All,” right? In the title, who would you say is the “us”? Most faculty are like, well, the “us” is the students. But is it also the faculty, the institutions, society at large? We’re so used to being the ones who deliver the knowledge to the students, but it seems like AI is going to teach more than the students. I don’t mean it is going to teach more than the students and replace us, but it is going to teach everyone.

Tim: Yeah. Everyone has a different lens. If a faculty member were to read that title, they’re going to say, “Oh, it’s going to teach me.” If a business CEO reads it, they have a different perception. But the real answer is everyone. It’s going to teach everyone.

In fact, I was recently reminded of something I said in a presentation back in, I think, March of 2023. A person came up and said, “You said something in your presentation. I wrote it down, put it on a sticky note on my monitor, and I share it with my students every semester.” I said, “Well, what was it?” He said, “You said that AI will expose the mediocrity of society.” I said, “Oh, I don’t even remember. I forgot I even said that.”

But when you look at it that way, AI will expose the mediocrity of society. When AI can do things better than us, it’s going to show us that the thing I’m doing here is pretty mediocre compared to what AI is now, or later will be, capable of. There are brilliant people all over the world, but even so, AI is going to become more brilliant.

I was invited to OpenAI headquarters, and I was very fortunate to have a conversation with the vice president of OpenAI. We were talking about AGI and things that AI can do. He told me a story that I thought was pretty cool. The researchers at OpenAI were working on a research project that they couldn’t figure out. There was a gap in the research that was stopping the project.

He gave their system 24 hours of compute time. When we do a prompt in ChatGPT, a few seconds later we get an answer. Well, this system was working on their problem for 24 hours straight. Throughout that process, it found a study from 1960 and related that 1960 study to their current problem. It solved the gap they were missing.

What was most remarkable, he said, was that their high-level researchers couldn’t find that 1960 study. He said that even if they had found it and read the research study, they would not have been able to make the correlation between that issue and this issue. But AI did, and it allowed them to keep moving forward.

He said, “What is AGI? At some things, we are already there. We have achieved AGI. At other things, we haven’t. It’s going to take more time.” But in either case, it’s progressing very rapidly. Things are changing, and they’re going to change even more as time goes on and these systems get better.

Subscribe now

The AI Literacy Gap: 97% vs. 20%

Anand: Tim, we really appreciate you taking the time to share with us and have this conversation, and also for exposing the mediocrity in so much of what’s going on in education. For anybody listening, where can they find out more about your work? Where can they find you online?

Tim: LinkedIn. I’m on LinkedIn. They can also find me at evolveaiinstitute.com. I think those are probably the two best places. My email address is accessible from either place.

Something I did want to share, too, and I had a little bit of a conversation with Stephan about it, is that when I was at the White House AI task force on education meeting, it was led by Melania. It included the secretary of labor, the secretary of education, the secretary of energy, CEOs from Microsoft, OpenAI, IBM, and all the huge corporations. They shared their initiatives for what they’re doing to contribute to AI and education. Everything was K-12.

The Google CEO said, “I’m going to donate this much money for the initiatives for K-12,” and so on. That meeting lasted somewhere around an hour. After the meeting, there was a three-hour reception. At this reception, I made it a point to talk to as many people in the room as I could. “Hey, what do you do? What do you do? What’s your plan?”

I found out very quickly that I was the only educator in the room. It was 100% politicians and CEOs. Once I realized that, when I was able to speak to some White House staff, I said, “You know what? This is really, really concerning to me because nobody knows the inside of a classroom better than teachers and faculty. Certainly, they know it better than politicians and CEOs, because they’re there doing it every day. So I really hope you’ll get more educators involved.”

I was shocked I was the only educator there. I couldn’t even believe it. I went to this White House meeting thinking it would be so cool because I would get to interact with other faculty members, find out what they’re doing, and learn from them.

There was one CEO who taught grade school 15 years ago. There was a White House leader who was involved with a college 10 years ago. That was it. That was the only connection.

The other thing I expressed to the White House staff that I found extremely concerning was that these White House initiatives are K-12, yet we have college students graduating and entering the workforce, and they don’t have these same initiatives. They are so strongly needed because students are going through two-year and four-year colleges being told, “You can’t use AI. You can’t do this.”

Then they get to the workforce, and they’re told, “You have to use it.” The story I love to tell is that the CEO of Shopify, which is a hundred-billion-dollar company, released an internal work memo. He didn’t release it, but it got leaked.

Once it got leaked, he went on X and said, “Hey, since this was leaked, I’ll share it anyway.” The two things that really stuck out to me in this memo were, number one, using AI on the job is an expectation. He went on to say that employees using it are 10 to 100 times more productive than those who are not. Number two, before you hire someone new, you need to justify why AI cannot do the job better.

So you have Shopify and other companies in other industries doing similar things, yet we have students we tell, “You can’t touch AI. It’s banned.” Now, with this White House initiative, we have students starting in kindergarten learning AI literacy, and they go K-12. Then they get to college and they’re told, “You can’t touch it.” They’re like, “What world is this? This is just craziness.”

Then they progress into industry, and they’re told, “You have to use it. You have to use this AI tool embedded into Excel to become more productive.” If they say, “I don’t know how to do that,” how do they even get through the interview? How do they maintain the job if they’re not on the same level as their competitors?

So I just think it’s so important that faculty really take a close look at how they can contribute to the success of their students in the future they’re entering. If they don’t, it’s going to be really hard on our students, and that’s the number one priority: our students.

Subscribe now

Why One AI Course Isn’t Enough

Anand: You put that so well, Tim. That’s great. Thanks for leading that charge. You’re right. We need to get more faculty at the table, and we need to make sure our colleagues are coming forward to figure out what’s best for the students. That’s really what it’s about. All right. Thank you so much, Tim. This was great. Thank you. I really appreciate it.

Tim: I appreciate the invitation. And let me just add, Stephan, I’ve been following your work since day one, when I saw you on LinkedIn. I’ve always had high respect for your writings and the things you’re doing. So being here is a great honor.

Anand: I appreciate it. It was great to meet you and see everything you’re doing as well. Thanks.

Subscribe now

<Five months ago, I argued against the Presiden t's $4 trillion tariffs at the Supreme Court.

In 237 years, the Court had never struck down a sitting President's signature initiative. Legal scholars said it was impossible. Some of my own colleagues said it was impossible.

We won. 6-3.

But the real story isn't what happened in that courtroom. It's what happened in the months before. And its the subject of my TED talk.

I had the best legal team in the nation, especially Colleen Roh Sinzdak, the most outstanding legal strategist I know. Huge thanks, too, go to the Liberty Justice Center (and in particular its fearless and hyper-intelligent leader Sara Albrecht), who organized the client small businesses, as well as to the brave small businesses themselves.

I also had four teachers preparing me.
A mindset coach who'd worked with Andre Agassi.
An improv coach who taught me that "Yes, and" works in Supreme Court arguments the same way it works everywhere else.
A meditation coach who taught me stillness.
And Harvey.

Harvey predicted many of the questions the Justices asked — sometimes almost word for word. Brilliant. Tireless. Occasionally insufferable.

Here's the catch: Harvey isn't a person.

Harvey is a bespoke AI I built over the last year with a legal AI company, trained on every question every Justice has asked in oral argument for 25 years, and everything they've ever written.

AI can predict. AI can analyze. What AI cannot do is the one thing that actually won the argument.

Connect. Read the room. Hear not just a Justice's words, but her worry — and answer the worry.

That is the irreducibly human skill.
Find yours. Go deeper. In this age of AI, that's where your edge lives.

The talk is live.

Subscribe now

Today, many schools are wondering what they can do to support students as we enter an AI World/Economy.

One relatively simple thing is to support speech & debate. It’s something any school can do.

Subscribe now

If you are able, you can also support speech & debate directly, just as Neal Katyal and Harvey are doing.

Subscribe now

Yesterday, I read this Axios piece by Dan Primack —”AI creates a mess for private equity”—and it sent me down a rabbit hole I haven’t been able to climb out of.

Primack’s argument is straightforward: Private equity, which is a medium-term asset class that typically holds portfolio companies for 4-7 years and then exits the investment, is struggling because AI is making it too difficult to predict what the economic value of any line of business, let alone a specific company, might be worth in 4-7 years. Basically, when a PE firm buys a company today and tries to project what someone will pay for it in 2030+, they are—in the words of one veteran Primack quoted—throwing at a dartboard blindfolded.

Read that again. The most sophisticated capital allocators on earth—people who employ armies of analysts, who can pay for any consultant, who have direct relationships with the AI labs themselves—are openly admitting they cannot model what the world looks like in 4 years.

And then I thought about the sixteen-year-old in high school

Subscribe now

The bet we ask teenagers to make

Currently, in May 2026, a junior in high school is being asked to seriously consider what they want to major in. They’ll apply to college next fall. They’ll start their freshman year in 2027. They’ll graduate, if all goes well, in 2031.

That’s a five-year forecast. One year longer than the PE hold period that the industry is now describing as a “material weakness.” Two years longer than the window Primack’s sources call un-modelable.

We are asking sixteen-year-olds to make a more confident forecast about the labor market than Blackstone is willing to make about its own portfolio.

And it gets worse, because PE has tools the student doesn’t:

• A PE firm holds twenty-plus companies. A student bets on one major.

• A PE firm can write down a loss and redeploy capital. A student who picks wrong has spent four years and—if they borrowed—six figures of debt on a credential the market has discounted.

• A PE firm marks to market quarterly. Universities don’t mark anything. The accounting major whose graduates can’t get jobs in 2029 will keep enrolling students in 2030.

• A PE firm has analysts covering frontier AI capabilities full-time. The student has a guidance counselor whose advice infrastructure was calibrated to the 1990s labor market.

The PE industry, looking at a three-and-a-half-year horizon, is openly telling the financial press that the modeling has become impossible. The education industry, looking at a five-to-six-year horizon, is still handing out major declaration forms and pretending the math works.

Subscribe now

Why this is worse than the PE problem

Primack frames the PE issue as a temporary embarrassment that will be partly absorbed by Q1 markdowns and renegotiated debt. The industry’s long-term nature, he writes, “always one of its greatest strengths, is becoming a material weakness.”

The same sentence describes higher education almost perfectly. Substitute “credential” for “portfolio company,” “graduating cohort” for “exit multiple,” and “tenure system” for “fund structure,” and the analysis ports cleanly. The four-year degree—long sold as the patient, durable bet—is the educational equivalent of the seven-year hold. Both rest on the assumption that you can lock in capital today against a stable future. Both assumptions are eroding for the same underlying reason.

But the cost falls in a different place.

When PE gets the model wrong, the loss is borne by limited partners—pension funds, endowments, sovereign wealth funds. These are diversified institutions. The pension can absorb a bad vintage. The endowment can offset a write-down. The sovereign wealth fund can wait it out.

When a student gets the bet wrong, there is no portfolio. There is one life, four years, the opportunity cost of those years, and—often—debt that follows them for decades. The “exit multiple” they’re trying to model is their own first job. There is no diversification. There is no patient capital cushion. There is just the kid.

Subscribe now

The advice infrastructure is broken

Here’s what I keep coming back to. The people advising students—guidance counselors, parents, college admissions officers, the BLS Occupational Outlook Handbook, the salary projection websites, the “best majors for 2026” articles—are extrapolating from historical patterns. When a parent tells their kid to go into accounting because it’s stable, they’re drawing on forty years of accountants having stable careers. They aren’t pricing in that the median accounting task—reconciliations, tax preparation, audit support, financial statement analysis—is precisely what current AI does best.

The advice sounds prudent. The advice may be catastrophic. And the people giving it have no way to know which.

This isn’t a failure of intelligence or care on the part of counselors and parents. It’s a structural problem. The advice infrastructure was built for a slow-changing world. The world is no longer slow-changing. Even Primack’s PE veterans, with all their resources, can’t see three and a half years out. A guidance counselor with a caseload of one hundred students cannot be expected to do what Blackstone cannot do.

Subscribe now

The institutional response will be too slow

Here’s the part that should make every parent of a high schooler pay attention. Primack’s piece notes that PE will continue making investments anyway, because limited partners keep allocating dry powder despite the DPI drought. The capital has to go somewhere. The funds were raised. The fees must be earned.

Universities are in the same position, only more so. Tenure, accreditation cycles, facilities debt, administrative overhead, and the four-year program structure all create momentum that prevents adaptation. A computer science department whose graduates struggle to find work in 2029 will not shut down. It will keep enrolling students in 2030, 2031, 2032, because the institutional machinery requires it. The university essentially cannot acknowledge in real-time that its product may be losing value, because acknowledging it would collapse the enrollment that funds the institution.

PE will probably take five to ten years to fully reckon with the modeling problem, because the funds raised in 2023–2025 still need to be deployed. Education will take fifteen to twenty years, because tenure and accreditation move slower than capital markets. Which means the students currently in high school will make their decisions inside an institutional framework that will not adjust in time to help them. The advice they get will be wrong. The credentials they pursue may be devalued. And the institutions providing those credentials will keep operating as if nothing has changed.

My gripe with the current educational system

Here’s what actually keeps me up at night. It isn’t that we don’t know which professions will survive AI. Nobody knows that. The PE veterans don’t know it. Dario Amodei doesn’t know it. Demis Hassabis doesn’t know it. Not knowing isn’t the problem.

The problem is that the educational system refuses to admit it doesn’t know.

Walk into almost any high school in America, and you’ll find the same machinery still running: course catalogs organized around traditional career pathways, guidance counselors steering students toward “stable” professions, AP tracks designed to feed pre-med and pre-law and pre-business pipelines, college admissions essays asking sixteen-year-olds to articulate “what you want to be.” The whole apparatus is calibrated as if we’re confidently sending students down particular career paths toward known destinations. We’re not. We’re sending them toward destinations that may not exist when they arrive, and the system is pretending otherwise.

Look at where AI has already gone. It came for the coders first—and that one stings, because coding was supposed to be the safe bet, the future-proof pathway, the thing every guidance counselor in America was telling kids to pursue five years ago. Now it’s coming for financial services. It’s well into the legal profession—document review, contract analysis, legal research, the work that used to fill the first three years of a young attorney’s career. It’s making serious inroads into research itself, which is a strange thing to watch, because research is what professors do, and professors are the ones training the next generation for the professions AI is now eating. It’s pushing into medicine—diagnostic imaging, differential diagnosis, treatment planning, the cognitive core of what doctors actually do.

And AI is just getting started. We are, at most, four years into the generative AI era. The capabilities curve is still bending sharply upward. Anyone who tells you with confidence which professions will be intact in 2031 is doing the same thing Primack’s PE sources are doing—except without the self-awareness to admit they’re guessing.

We cannot, with a straight face, tell students that any profession is sacred. We cannot promise them that medicine will look like medicine, or that law will look like law, or that finance will look like finance. We don’t know. Some will survive in recognizable form. Some will be transformed beyond recognition. Some will mostly disappear. We don’t know which is which, and pretending we do is a kind of malpractice.

But here’s the thing—there is one thing we do know. We know that the fundamental human capacities will matter no matter which professions emerge on the other side. Real-time reasoning. Argumentation. Evidence evaluation. Ethical judgment under pressure. The ability to defend an idea when an intelligent person is pushing back across the table. The ability to change your mind when the pushback is good. Dialogue across genuine disagreement. These are the things that don’t get commoditized when the labor market gets reshuffled. These are the things that matter in every possible future configuration, because they’re the capacities that let a human direct AI rather than be directed by it.

This is the one thing we know. And what does the educational system do with the one thing we know?

It often treats it as an extracurricular. It funds it as a club. Speech and debate—the single most direct training ground for exactly the capacities that will survive AI disruption—is offered at most schools as an after-school activity for the kids who happen to be interested, run by a teacher who is doing it on top of a full or nearly full course load, and a limited budget. Meanwhile, the resources flow to the coursework training students for the careers we cannot promise will exist.

That is my gripe. We are lowballing the only thing we know is durable, while we keep dressing up the things we don’t know in the costume of certainty. We treat the AI-resilient capacities as enrichment for the interested, when they should be the currency of the emerging AGI world. Every student should be doing this work. Every student. Not as a club. As the core.

The PE industry is at least having the conversation. They’re saying out loud that the modeling is broken, that the long-term bets they’ve always made don’t pencil out the way they used to. The educational system isn’t even at that stage. It’s still printing the same course catalogs, still routing students into the same pipelines, still telling sixteen-year-olds to pick a major and plan a career as if the next decade will resemble the last one.

It won’t. And we owe these kids better than the pretense that it will.

Subscribe now

So what do we tell the sixteen-year-old?

I don’t think the answer is “don’t go to college” or “skip the major.” That’s the wrong frame, and it’s not what I’d tell my own students. The answer is that the entire pick-a-major-as-life-bet model is itself the obsolete asset. It’s not that students are picking the wrong majors. It’s that the framework of treating a major as a five-year forecast on labor market value was always fragile, and AI has now exposed the fragility.

The replacement isn’t a smarter major. It’s a different theory of what education is for.

If you cannot reliably forecast which specific knowledge bundles will be valuable in 2031, you have to bet on capacities that survive whatever specific economic configuration emerges. Real-time reasoning under pressure. Evidence evaluation. Adversarial argumentation with a human across the table. Ethical judgment when no rule covers the case. The ability to defend an idea under cross-examination and to change your mind when the cross-examination is good. Dialogue across genuine disagreement.

These are the things I’ve been writing about as the D.E.B.A.T.E. framework, and I keep coming back to them, not because debate is my hammer and everything looks like a nail, but because these capacities are domain-independent. They don’t depend on guessing right about which industries will exist in 2031. A student who can think well, evaluate evidence, and defend ideas under cross-examination has a portable capacity that survives any specific economic configuration. A student who has memorized the 2026 marketing curriculum does not.

The Axios piece is, on its face, about private equity. But the deeper story is about every institution that asks people to lock in capital today against a future that nobody can model. PE is reckoning with this in real time. So is higher education—it just doesn’t know it yet, or it knows and can’t say so out loud.

Subscribe now

Capabilities are accelerating. OpenAI’s Greg Brockman says they’re ~80% of the way to AGI; Grok 4.3 ships cheaper inference and custom voice cloning. China’s open-weight MiMo 2.5 Pro matches Opus 4.6 on coding benchmarks using 40–60% fewer tokens — Scott Galloway’s framing: China’s play is to flood the market with cheap open-weight models rather than chase closed-frontier parity. An Anthropic co-founder is now targeting full recursive self-improvement around 2028 (the AI 2027 paper had previously stated 2027).

Government is stepping in. Microsoft, Google, and xAI will give the U.S. government early access to frontier models for pre-release safety review — a major shift, and one that raises unresolved questions about what “security” review actually covers (capabilities? ideas? alignment?).

AI pluralism: Different models give different answers to the same ethical dilemmas — feature, not bug. The risk is monoculture if one lab wins AGI first. There’s no single “human value alignment” to align to; the better path is multiple frontier models checking each other.

Deployment & integration. Anthropic (with Blackstone/Goldman) and OpenAI (”the Deployment Company”) are launching mid-market services arms — pre-IPO moves to capture integration value, not just frontier capability. DoW signed agreements with most major labs (Anthropic notably absent; Susie Wiles reportedly trying to repair that). Microsoft is putting a legal contract agent in Word. UAE plans to run 50% of government on agentic AI within two years. Suno is at a possible $5B valuation.

Robotics. China deployed humanoid robots over the Mayday holiday — partly norm-building. Meta bought a robotics company. U.S. firms are scaling production, but mostly for factories/warehouses first.

Consciousness & jobs — the real friction. Anthropic and Google have hired consciousness researchers. Richard Dawkins called a model conscious and got dogpiled. Coinbase and PayPal both announced AI-driven layoffs the day of recording. Altman is cooling on UBI; meanwhile a Chinese court ruled companies can’t fire workers just to replace them with AI. Reese Witherspoon got attacked for telling her book club to learn AI — illustrating the U.S. backlash dynamic heading into the midterms.

The hosts’ closing point: People aren’t unready for the technology — they’re unready for the moral dilemmas the technology forces (authorship, representation, misinformation, consciousness, displacement). Without pluralistic, non-polemical public debate, the discussions get captured by political fear and stop producing solutions.

Subscribe now

The New Yorker has a wonderful new article on high school debate.

Throughout the pieces, you see —

Students learning
Students developing “durable/soft skills
Students developing *agency*
*Students using *devices* in productive ways

*Phones don't have to be confiscated
*We don't have to try to put AI in a Yondr Pouch

Dewey's back.

We already know how to redesign teaching & learning for the A(G)I World/Economy.

Subscribe now