[Our classes on Claude are starting this week, and the first video recordings are available to those who signed-up. We do have a free recording on Claude Skills available]
At the end of every school year, I try to take stock. This year feels a little harder than usual because the questions for most have changed.
For three years, the conversation about AI in education for most people (a few of us pushed much further, much, much earlier) was largely about using AI in the classroom — a tool here, a chatbot there, some traffic lights, and a policy somewhere.
Now, more people (unfortunately, not everyone) have moved past that. We’re now arguing about instructional design, and beyond that about the structure of schooling itself. That’s the right altitude, and those are really the only thing that matter.
So let me say plainly what this post is and isn’t. This is not a piece about AI in the classroom. It is a piece about what a changing world asks of schooling. The myopia I want to avoid is the one that treats “AI in education” as a sector problem — a question of which apps to adopt and how to stop students cheating with them — when the actual subject is the world these students are being educated into: a labor market losing its bottom rung, a geopolitics rearranged around compute, a reckoning over what human worth even rests on, an information environment with no firm floor, machines that are starting to move and act.
None of that fits neatly inside a chatbot or a school AI policy, and none of it can be cordoned off from the question of what school is for. Everything here is interrelated, which is exactly why the sector-only view misses it. As the technology keeps developing, the honest conclusion is that we are entering something like a new industrial era, and our institutions are only beginning to notice. Here is where my thinking stands as the year closes.
1. Learning is cooked without instructional redesign
“I think if we continue to teach and evaluate students as if we were in a pre-AGI world, it’s not going to work and it is going to lead to atrophy of learning how to think… there are a lot of other things where we should just totally teach — totally change how we teach, or how we learn or how we evaluate. And if we don’t do that, I think there will be significant atrophy in people’s critical thinking skills.”
— Sam Altman, CEO, Open AI @ Stanford CS 153, 2026
Let me say the quiet part plainly: the AI can do the homework. Not approximate it, not assist with it — do it, well enough to earn the grade, in a fraction of the time. Every approach that pretends otherwise — the honor code, the detector, the stern reminder — is a wall built against water. Students will use it because it works, and prohibition is not going to change that. The only live question left is what we ask them to do instead.
We now have population-scale evidence of what happens when the answer is “nothing different.” In The Generative AI Learning Penalty, Strömberg, Lei, and Wu tracked 26,811 Chinese students across grades 7–12 for thirty months, using the staggered way AI tools reached different students to separate cause from coincidence. The pattern is the exact trap you would fear. AI adoption raised homework scores by 18% and cut homework time by 30% — the assignments got easier and faster. But monthly exam scores fell 20% within six months, and the high-stakes entrance-exam scores, the ones that actually decide a student’s future, fell 18 and 24%, with the full penalty taking about two years to surface. The work looked better while the learning got worse, and the bill came due slowly enough that no one would catch it inside a single semester. (This is not one anomaly from one country: a separate decade-long panel of millions of U.S. math interactions found the same fork — faster completion, less learning — the moment chatbots arrived.) A UC Berkeley study of more than 500,000 grades at a large, selective U.S. university adds the grade-level version of the same fingerprint: in writing- and coding-heavy courses, grades climbed sharply after ChatGPT — and the jump showed up in homework, not exams, which is what outsourcing, not learning, looks like.
Here is the finding that should reorganize how we think about an assignment. The damage was not spread evenly. It was concentrated in the roughly 80% of AI users whose behavior fit outsourcing — exceptionally fast completion paired with high scores, the fingerprint of a task handed to a machine. The students who kept working at about the same pace as their non-AI peers barely lost anything at all. The learning penalty, in other words, is not a property of the technology. It is a property of the assignment. Give students work that can be outsourced and most of them will outsource it and learn less; the tool simply exposes a hollowness that was always there. And the losses ran deepest in social science, then STEM, then languages — and worst of all for the youngest students, the highest achievers, and boys, which is to say the penalty falls hardest on exactly the kids we reassure ourselves are fine.
It is tempting to treat the standardized, scope-and-sequence school as the stable thing AI is disrupting from the outside. I’ve come to think the truth runs the other way: a system that produces standardized minds — shaped by the same sequence, graded by the same rubric, certified by the same credential — is manufacturing exactly the population AI is best at substituting for. The factory model isn’t a hedge against the machine. It is the thing the machine most directly replaces.
Seen that way, the panic that AI is “killing critical thinking” looks misdirected. We built a curriculum around information retrieval and then ranked children by who could retrieve fastest; our two decisive instruments, the timed test and the write-it-alone paper, mostly reward retrieval and pattern-matching against an answer key. AI isn’t destroying the thing we claimed to prize. It is dissolving the three things the system actually delivered — a way to rank students, a scarcity worth competing over, and the right to teach the same units the same way every year. What’s breaking isn’t thinking. It’s an easy sorting function that has stopped sorting. This is starting to be said plainly: that schools keep treating AI as a cheating problem instead of the civilization-level shift in how people learn, work, and compete that it actually is.
The most radical version of this argument goes after the fundamentals themselves. Tim Dasey asks whether the “three R’s” are still fundamental in a world where AI can read, write, and calculate on demand: if literacy and numeracy were always a vehicle to higher-order thinking rather than the thing itself, then perhaps they become specialized skills rather than universal ones, and the years now spent drilling them could go to the judgment and durable skills the machine can’t hand you. I don’t think the case is settled — the claim is genuinely uncomfortable, and Dasey concedes he doesn’t know the answer — but it is the right kind of question, because it refuses to treat the curriculum as the one fixed point while everything around it moves.
The boldest version of this is now running at national scale, and it shows why removing the tool is not the same as solving the problem. This month Norway — having already pulled smartphones out of its classrooms — moved to bar generative AI for children in grades one through seven, ages six to thirteen, effective when classes resume in late August, permitting it for fourteen-to-sixteen-year-olds only under a teacher’s direct supervision and pairing the restriction with legislation to put physical books back in classrooms and reverse the tablet rollout the country embraced after the iPad arrived. Prime Minister Jonas Gahr Støre’s rationale is the developmental one the learning-penalty data points to: that leaning on the tool lets young students skip the essential steps by which a person first learns to read, write, and do arithmetic at all — and in the early acquisition years, where those skills are still being laid down and the outsourcing penalty fell hardest, walling the tool out has a real case. But here is the thing the policy quietly assumes: that the assignments waiting on the other side of the ban are still worth doing as written. A ban and a bolted-on chatbot are two faces of the same error — both change what students may touch while leaving the instructional design exactly where it was. If Norway protects the foundational years and then hands its fourteen-year-olds the same outsourceable worksheets and take-home essays it always assigned, it will have bought time and changed nothing; the moment the tool is permitted, the old work fails the same way it fails everywhere else. Prohibition without redesign doesn’t beat the problem. It postpones it — and integration without redesign doesn’t beat it either. The variable that decides whether either approach works is never the tool; it is whether the work itself was rebuilt to be worth doing.
And the positive case is not merely theoretical. In a five-month randomized trial across ten Taipei high schools teaching Python, a Wharton-led team found that an AI tutor built to proactively sequence practice rather than reactively answer questions raised unassisted final-exam scores by 0.15 standard deviations — by one estimate six to nine months of additional schooling — with the gains flowing through higher engagement. Read alongside the penalty studies, the lesson runs the same way from the other direction: the technology is neutral, and the design decides. Pointed so students can skip the work, it hollows out learning; built to guide the work, the very same models measurably deepen it. The reactive chatbot most schools rolled out is the weak version of this — and the distance between it and a tutor engineered to teach is exactly the distance between bolting AI on and designing with it.
So the move is not to fight the tool. It is to redesign the work so that doing it and learning from it become the same act again — work you cannot hand to a machine and still collect the credit. Make students argue a case live and hold it under cross-examination. Make them build something and ship it to real users. Put the assessment inside the performance, the design problem, the venture, the debate round — the thing done in front of other people, in real time, where the doing is the point and the doing is visible. That is what performance-based assessment, design thinking, and entrepreneurship actually have in common, and it is why they survive contact with AI.
And none of this is hand-waving; teachers and school leaders have been building these alternatives for years. Jerry Crisci has argued for instructional redesign at Scarsdale’s Center for Innovation for the better part of a decade; Sabba Quidwai puts design thinking at the center of her Designing Schools: How Design Thinking Makes YOU Irreplaceable in the Age of AI; Phillip Alcock has rebuilt project-based learning into a “PBL 2.0” that runs with AI rather than against it; Jason Gulya has rebuilt writing instruction around the process rather than the product, swapping the final paper for portfolios in which students document, self-report, and defend each step of how they actually worked, AI use included. Daveed Gartenstein-Ross’s Expert Theory turns a subject into a live, AI-generated simulation — now the backbone of courses at Duke and Arizona State — where students act inside the problem instead of writing about it, and the cheating worry dissolves because the game demands more thinking than a paper ever asked; and the case Alan Coverstone, Anand Rao, and I have made for academic debate treats the live, contested round as the one assessment a model cannot sit on a student’s behalf. The redesign is not a theory waiting to be invented; it is already being practiced, just not yet at scale. You can even see it built into a building: at St. Vrain Valley Schools in Colorado, Joe McBreen, the assistant superintendent for innovation, runs a 50,000-square-foot Innovation Center where the work is the assessment — students earn industry certifications and drone-pilot licenses, run capstone projects and internships with real industry partners, and present their AI work at a public expo with live demos, a walk-through of their design process, and an ethical reflection. It is a public district, not a boutique microschool — an existence proof that “put the assessment inside the performance” can be the operating model of an actual school rather than a slogan, and that the redesign is not reserved for the well-funded fringe.
Practitioners are increasingly blunt about the corollary — that the take-home written essay, in particular, has outlived its run as a trustworthy measure of a student’s own thinking. Schools that make this shift keep teaching. Schools that keep writing policies and hoping the homework still measures something are, by the most rigorous evidence we have, watching their students learn less with every year they wait.
There is a clean way to name what has changed. The University of Malta’s Matthew Montebello argues that assessment used to be a structural mechanism that quietly enforced cognitive engagement — you could not produce the essay without doing the thinking — and that generative AI has severed that link, moving cognition “from cognitive necessity to cognitive choice.” The result is what he calls a cognitive engagement gap: a finished assignment no longer reliably signals that any learning happened, which is precisely what the grade-inflation and learning-penalty studies measure. His reframing of the design question is the sharp one — not “can AI do this task,” but whether the task keeps the use of AI “cognitively generative rather than cognitively bypassing.”
It is worth naming the upside directly: this is cognitive reclamation. The learning penalty is really cognitive offloading gone wrong — hand the thinking to the machine and the capacity quietly wastes, the “atrophy” Altman warns of at the top of this section. France Hoang puts the distinction in a line — AI is about productivity, but learning is about productive struggle — and the trouble is that the model is engineered to remove the very struggle that does the teaching. Redesigned work runs that in reverse. When the assessment lives inside a live argument held under cross-examination, a build shipped to real users, or a design defended out loud, the student has to do the analysis, the synthesis, and the judgment in real time and in front of others, because there is no seam left where a machine can quietly do it instead. The aim is not to keep AI out of the room; it is to reclaim, on purpose, the thinking that outsourcing had ceded — and, used well, the same tool accelerates the reclaiming, a model that interrogates a student’s reasoning or stages a harder counterargument, putting cognitive load back on the learner rather than lifting it off.
2. The school model itself is fragmenting
Redesigning the work assumes the institution around it will hold still. It won’t. For most of living memory there was essentially one model of school. From the 1970s through the 2000s, private schools were genuinely different from public ones — smaller classes, a religious affiliation, less standardized testing, selective admissions — but structurally they were near-identical. You could sit for the same AP and IB exams in either. Many private schools still administered the state tests. Different, but the same skeleton.
That skeleton is now cracking. Homeschooling, which spiked during the pandemic, never snapped back; it sits at roughly 3.4 million K–12 students, about 6.3% of the school-age population, growing close to 5% a year — roughly triple the pre-pandemic rate, with more than a third of reporting states at all-time highs. Microschools have gone from a curiosity to a sector; the National Microschooling Center now counts something like 95,000 microschools and pods serving more than a million students. And networks like Alpha School are expanding aggressively — roughly a dozen new campuses planned through 2026–27 — on a model that compresses core academics into a short “2 Hour Learning” block run by software and human “guides,” then hands the rest of the day to entrepreneurship, debate, robotics, and other deeper-learning work.
The impulse to rethink the architecture reaches inside the public system too. Kip Glazer, a Silicon Valley high-school principal, says that if she had a free hand she would scrap the grade-level K-12 structure entirely for something closer to a non-age-banded “monastery model.” Her analogy is sports: we already let a thirteen-year-old who can outrun the eighteen-year-olds play varsity, accepting and even celebrating human variability in physical prowess — yet in academics we still insist on sorting children by birth year and moving them in lockstep. Age-banding, on that view, is not a law of learning but an administrative convenience, and one of the first things a serious redesign would put back on the table.
You don’t have to admire any particular one of these to see the pattern. For the first time in decades, the architecture of the school day — not just its branding — is genuinely up for grabs. That is exactly what you’d expect at the start of a new industrial age.
3. We keep optimizing the part of school that should disappear
You might think a moment this open would push us to rethink the core task. Instead, we keep refining it. Once you grant that content is now a free utility, the strangest feature of an ordinary school day comes into focus: we are still spending our scarcest resource — a human being — on the one task a machine now does better. Muhammad Khan, an associate professor at NUST in Islamabad, put it plainly this year: a human teacher who spends most of class delivering content can feel like “a waste of time for both teachers and students,” and yet the whole system is still organized around the premise that to teach is mostly to deliver. He is right, and the waste runs both ways. The machine is endlessly patient, available at 2 a.m., and can tailor the explanation in a way no teacher with thirty kids ever could. Set against that, a human reading the slides aloud is not merely inefficient; it is a misuse of the most valuable thing in the room.
The trap is that the obvious response — bring in technology — has mostly meant aiming that technology at the same obsolete task — a pre-AI curriculum running essentially unchanged in an AI world. As Jason Gulya, who teaches writing at Berkeley College, has noted, the EdTech he encounters mostly extends “practices that we should be questioning, not extending.” That is the deeper failure beneath the shallow engagement: handed a machine that delivers content better than any human, we answered by building tools to make humans and software deliver content faster. We are automating the part of school that ought to be shrinking. The work to do first is deciding what world we are educating students for, building the urgency to change, and then redesigning instruction around it — not buying a faster version of the thing that no longer works. The reason to keep a person in the room cannot be to do what the machine does; it has to be the work the machine cannot touch — running the argument, pressing on the reasoning, coaching the performance, judging the live attempt, knowing one particular student well enough to push at the right moment. That is a different job than the one most teachers were hired, trained, and scheduled to do — and pretending otherwise wastes the teacher as surely as it wastes the student.
4. Schools are engaging — unevenly, and mostly at the surface
“When ChatGPT launched, I was like, yeah, we’re going to have one year of students cheating and not learning that much, and then the educational system is just going to redesign itself, and we’re going to teach people so much better. … And, honestly, I struggle to point to any significant systemic change that I’ve seen in the education system at large in the three and a half years since ChatGPT launched, and that was a prediction error for me. I thought that would have happened.”
— Sam Altman, Stanford CS 153, 2026
To be fair, there is more effort than there was. But it falls along a steep spectrum. At one end are schools genuinely working on instructional design and weaving AI into the curriculum. In the middle are the many schools rolling out a chatbot and encouraging teachers and students to use it. Then there are the schools that have written a policy and little else — which, to me, is nearly indistinguishable from doing nothing — and finally the ones that have simply ignored the whole thing.
The data show how shallow the center of that distribution is. By 2025, 54% of students and the share of K-12 teachers using generative AI for work doubled in a single year, from 25% to 53%; in Europe the saturation is starker still, with an OECD/EU framework reporting that 88% of 13-to-15-year-olds and 96% of 16-to-18-year-olds use AI tools for schoolwork at least weekly and that 16-to-24-year-olds use generative AI at nearly twice the rate of the general population. RAND’s nationally representative panel caught the same climb in real time: students using AI for homework jumped from 48% to 62% across 2025. And students draw their own moral line — nearly 80% say using AI to understand an assignment isn’t cheating, but only about 45% say the same about using it for answers — which is the very distinction between learning and outsourcing that the first observation turns on. But only about 45% of principals report any AI policy, just over a third of teachers report an academic-integrity policy — Gallup found fewer than one in ten get formal guidance on any specific AI task — only about 35% of district leaders offer students any AI training, and more than 80% of students say a teacher never actually taught them how to use AI for schoolwork. Some of the energy that does exist is misdirected — reacting to AI by bolting on more STEM, which isn’t really the point. The bulk of the effort is a thin layer of basic literacy and a posted policy. Carlo Iacono has a sharper name for that shallow middle — the great performance: institutions caught in a “middle ground of doom,” dabbling and hedging, doing just enough to call themselves AI-aware while quietly hoping the whole thing blows over, wanting the productivity gains without the identity crisis that real change would force. It won’t blow over. That pattern now has the force of law: states moved hard on AI in education in 2026 — the PIE Network tracked nearly a hundred K-12 bills, part of more than 1,500 AI-related bills introduced nationwide — but the new statutes cluster on guardrails (data privacy, parental opt-outs, bans on letting AI make high-stakes calls about students) and on AI-literacy or computer-science graduation mandates: the bolt-on-more-STEM reflex again, not a rethinking of what the years are for. The requirements are also hardening into law: Maryland’s 2026 AI Ready Schools Act now requires an AI coordinator in every district and compensated statewide training for teachers — what the instructional coach Jimmy Norman calls AI training “becoming the law, not a nice-to-have” — though here too the mandate runs to literacy and coordination, not a redesign of the work. There is very little curriculum reform, and almost nothing more radical than that — which is what will ultimately be needed.
The field’s own leadership is increasingly blunt about what that radical version is not. Seiji Isotani — a Penn faculty member, UNESCO Chair, and president of the AI in Education Society — argues that schools are repeating the mistake they made with computer labs, one-laptop-per-child, and the internet: obsessing over the technology instead of the learning. “Putting ChatGPT in a classroom is not transformation,” he writes; “buying licenses is not transformation; deploying another platform is not transformation.” The world long ago expanded access — more than a billion children are in school — without solving learning: only three in ten children in Brazil reach basic literacy by age ten, against nine in ten in high-income countries. His alternative, the “phygital school,” is less a place than an ecosystem — physical, digital, and social dimensions fused around a single objective, mastery learning for all, in which individual, collective, and artificial intelligence work together rather than one being bolted on top of the rest. It is the argument these notes keep making, raised to the level of the whole system: the transformation was never the tool; it is the redesign the tool is supposed to serve.
And the most common posture of all is simply to wait. Priten Soundar-Shah, whose Wiley book Ethical Ed Tech presses educators to lead rather than defer on AI and student safety, names the script schools recite to themselves: that they will “wait to see where this technology ends up,” or that a task force is “tracking the development to inform future policymaking.” His rejoinder is blunt — the pace is not going to settle and may only quicken, so the wait-and-see era “was never here”. What the moment calls for instead is a nimble, iterative strategy: asking the big questions now, supporting teachers substantially, and building the ethical scaffolding — privacy, equity, digital safety, the actual hard cases his book works through — before the next tool lands, rather than the “reactive bandaids” applied after it does. The committee convened to study the question is, on this view, just the wait-and-see posture in respectable dress.
Dasey names the failure mode precisely: the workshop-and-committee response — send a few people to a training, convene a committee, let the motivated few tinker — is not commensurate with the disruption, and “nothing accumulates,” because each gathering restarts from roughly the same place. His alternative is a useful picture of what serious engagement could look like: an “AI immersion” week in which a whole school community — teachers, students, parents, administrators — sets its normal business aside and works the hard questions together until everyone leaves with a real deliverable and a name attached to it, moved past the binary of “AI will ruin education” versus “AI will transform it” by building and deciding something rather than watching someone else do it.
5. Everyone is a builder now
The more radical thing is not exotic; it is already sitting on the desk. The same year the machine got good enough to do the homework, it got good enough to let students and teachers build. “Vibe coding” — describing software in plain language and letting the model write it — went from a curiosity to the Collins Dictionary word of the year in a single stretch. In a class I helped teach this fall, undergraduates were using these tools to build genuinely capable apps, and the ten-to-thirteen-year-olds I work with are already prototyping their own. A teacher can now build a custom tutor for a specific unit in an afternoon; a student can build the thing they wished existed. The shift is real enough to measure: in Anthropic’s analysis of roughly 400,000 coding sessions, people set the direction — about 70% of the planning decisions — while the model handled some 80% of the execution, and, strikingly, users from every major occupation succeeded at nearly the same rate as professional software engineers. The barrier was never the syntax; it was knowing what to build.
This is the half of the story the doom narrative misses. The tool that flattens a paper into fake fluency is the same tool that hands a kid the power to make software, simulations, and instruments that used to require a team and a budget. If we only ever ask students to consume what the machine produces, we waste it. The interesting move is to make them build with it.
It also reframes what “AI literacy” should mean. As Dasey argues, most literacy frameworks train students to evaluate AI’s outputs — to spot bias and hallucinations, which is exactly the receiver’s role the machine is now absorbing fastest. The durable skill is the harder opposite, what he calls critical doing: deciding what to build, whether AI belongs in the task at all, how to break the work apart and direct it, when a technically correct answer is wrong for the situation. Receive-only training prepares students for a world that no longer exists; everyone now has to teach and manage the machine, not just grade it. Officialdom is, slowly, catching up to this. The flagship OECD/EU “AILit” framework, released in 2026 to feed the PISA 2029 assessment, organizes AI literacy into four domains that run Engage → Create → Manage → Shape — and it culminates not in evaluating AI but in shaping it, pushing learners to “move beyond simply using existing AI systems” to improve them and to “divide work intentionally between humans and AI.” That is the build-and-direct posture, written into an international standard. The catch is the familiar one: the framework leads with “Engage,” the foundational evaluate-and-critique layer, and that is where almost all classroom practice still stops — the part schools find easiest to bolt on.
There is now a pedagogy built around exactly that human-AI relationship. Mairéad Pratschke, who chairs digital education at the University of Manchester, calls it generativism — “a symbiotic approach to teaching and learning with GenAI,” grounded in the principle of learning as a process rather than a product. Instead of treating the model as a content dispenser or a cheating threat, she maps it onto established learning-design frameworks as a genuine actor in the work — a collaborator, an analyst, a facilitator the student thinks with — so that learning becomes co-creation, co-facilitation, and co-assessment. It is the constructive mirror image of the assessment redesign in the first observation: that one rebuilds the task so a machine cannot do it unseen; this one rebuilds the task so the student learns by directing the machine in the open. Both refuse the only move that actually fails — pretending the machine isn’t in the room.
Amarda Shehu, a George Mason computer scientist who teaches an AI-literacy course open to any undergraduate, pushes the idea further into civic territory. The point of the course, she writes, is not to settle students’ views but to build the practice of forming and revising a position from evidence — “the right to change one’s mind,” she calls it, “is the only ground from which one earns the right to hold a position at all.” Having grown up under a regime that told people what to think, she refuses to let the classroom become “a recruiting station for my own views”; what she wants is a graduating class that has not “imported its positions from the comment threads” or let itself be sorted into doomer and boomer camps. That is AI literacy as civic agency — not only learning to direct the machine, but learning to think for yourself in an information environment the machine is busy flooding.
6. We keep underestimating how smart the machines are getting
A quick note on something I’ve taken for granted up to here. Everything so far presumes the machines are capable enough to force these changes — so it is worth turning, for the next several sections, to what these systems are actually becoming — how capable, how cheap, how geopolitical, how strange — before circling back to what it all asks of school.
Most of us still carry a quiet assumption that we are, in the end, smarter than the machine — sharper at the things that actually count. It is getting harder to defend. Set aside the chatbot that drafts an email and look at the frontier of what these systems now do.
In May 2026 an OpenAI reasoning model produced an original proof disproving a conjecture Paul Erdős posed in 1946 — the planar unit-distance problem, open for eighty years — and the proof was checked and published by a group of mathematicians, several of whom had publicly mocked the company’s earlier, overblown math claims. Days later Google DeepMind reported that its system had resolved nine open Erdős problems, two of them unsolved for more than fifty years. This is not the machine reciting results humans already had; it is, with the appropriate caveats, machines generating genuinely new mathematics. The caveats are real — an earlier 2025 boast that GPT-5 had “solved” ten Erdős problems fell apart when the “solutions” turned out to be buried in the existing literature, and these problems are an uneven measuring stick — but the May results are the real thing, and they are not one-offs.
And it is not only mathematics. In cybersecurity, Anthropic reports that a preview of its Mythos model turned up thousands of high-severity vulnerabilities, including in every major operating system and browser — the same capability that hardens critical systems when pointed one way and threatens them when pointed the other. And in medicine — on OpenAI’s own physician-built benchmarks, so read with that grain of salt — the company reports that its latest model’s answers to real health questions were rated higher than physician-written ones for accuracy, completeness, and even communication, with fewer missed red flags than the doctors. Note where that last one lands: not on abstract reasoning, but on bedside manner, the “human touch” we keep assuming is the safe ground.
The same capability keeps illuminating the buried and the broken. A self-taught engineer reports that Claude Code helped him crack Linear A, the Bronze Age Minoan script that resisted scholars for a century, arguing it encodes an extinct Semitic language — a claim now under review at Rutgers and Cambridge, so hold it loosely. And biology is being recompiled: Texas A&M researchers report they regrew bone, joint, and tendon after amputating a digit in mice with a simple two-step molecular nudge and no added stem cells, suggesting mammalian regeneration may be dormant rather than absent. And the modeling is scaling up: the Chan Zuckerberg Biohub’s $500-million “virtual biology” program is training AI “world models” of cells and releasing an open-source protein-design engine, with its founders now calling their original goal — to cure, prevent, or manage all disease by century’s end — “too conservative.” Discount the founder optimism; the point for this section is the direction, which is the same everywhere — the machine increasingly generating science, not just summarizing it.
The trend beneath the headlines deserves more attention than the headlines. METR, which measures how long a task an AI can carry out on its own, finds that the length of work a frontier agent can complete autonomously has been doubling about every seven months for six years — and lately closer to every four, carrying the best systems from tasks of a few seconds to tasks that take a human expert half a working day — by Amodei’s account, METR now clocks the newest models at around four hours of autonomous work at even odds. Set that beside Moore’s Law, which doubled every eighteen months and rebuilt the modern world over sixty years: at a seven-month doubling you reach the same number of doublings in a little over two decades. The transformation that last time unfolded across a lifetime is being compressed into less than a generation.
And the people who forecast this for a living are not relaxing. The median estimate for artificial general intelligence has collapsed — the lead author of the influential “AI 2027” scenario moved from a forecast of 2070 a few years ago to around 2027, and now lands on roughly 2030, with wide uncertainty. Notice what that walk-back actually concedes: even the correction lands on “within years,” not “within decades.” Meanwhile the scenario’s structural calls — agentic AI in 2026, AI-driven layoffs, an arms race over compute and model access, autonomous weapons — are tracking with uncomfortable precision, in several places ahead of schedule, as much of the rest of these notes attests. And the people building these systems are saying it without much hedging: Anthropic’s Jack Clark expects that by the end of 2028 it will be more likely than not that you can tell a model to “make a better version of yourself,” and Alibaba’s Eddie Wu now calls human-level AGI “a certainty” — and only the start of the road to superintelligence. The deployment side is moving just as fast: Qualcomm’s Cristiano Amon calls 2026 “the year of agents,” software that no longer waits for a prompt but takes actions on its own. And Demis Hassabis, who a year ago put AGI “more than five years” away, now says we are standing in the “foothills of the singularity”, expects it around 2030 or possibly 2029, and calls the current wave of agents a “practice run” — a societal stress test for the far more powerful systems still to come; his boss Sundar Pichai puts AGI on the “closer side” of a three-to-five-year horizon and warns that even if it takes ten years, the technology three years out will be transformative enough that no one should wait to prepare. Reasonable people still argue about exactly when “superintelligence” arrives, and the labs have every reason to hype. But the burden of proof has flipped. The comfortable belief that human intelligence sits safely above the machine is now the claim that needs defending — and it is getting thinner every quarter.
That underestimation has a history the field is now reckoning with. Sam Altman, pushing back on skeptics like Yann LeCun — who has called LLMs a dead end — argues that “a whole generation of researchers held the field back” by being too confident about what scaling could not do, that betting against it now “feels quite misguided,” and points to the conjecture-disproving result above as proof that “LLMs are capable of figuring out new knowledge.” Read past the salesmanship and the pattern is the one this section keeps hitting: the careful, credentialed reasons the machine can’t do X have a short shelf life, and the experts who stake their identity on them keep getting overtaken.
Still, it is worth being precise about what is actually being predicted, because much of the disagreement is about words, not facts. “AGI” has no agreed definition, and the distance between definitions is enormous: the same 2023 survey of AI researchers that dated “high-level machine intelligence” to a median of 2047 put “full automation of labor” at 2116 — a sixty-nine-year spread from wording alone. Hold the definition fixed and the camps separate cleanly. Lab insiders, forecasting “powerful AI” or economically-defined automation, cluster around 2026–2030; the broad machine-learning community lands decades later — that survey’s median for human-level AI fell to 2047 (down from 2060 the year before, compressing, but still far beyond the labs’ horizon), and the 2025 AAAI panel found 76% of researchers doubt that scaling today’s methods will get there at all; the betting markets split the difference, landing around 2030–2033. So the honest summary is not a date but a crux — whether today’s paradigm of scaling plus reasoning plus agents is sufficient, or whether a breakthrough still stands in the way. That, not the calendar, is the real disagreement.
And there is a faster variable still, the one the labs are now turning their own systems toward: using AI to improve AI. The idea is old — in 1965 the mathematician I. J. Good argued that a machine able to design better machines would touch off an intelligence explosion that leaves human intellect far behind — and recursive self-improvement is the engine beneath both Clark’s prediction and the AI 2027 scenario. What’s changed is that the loop is no longer purely theoretical: Dario Amodei reports that coding agents already deliver 20-to-40% gains in software development, with some of his own engineers no longer writing code by hand and the gap between major model releases falling from many months to weeks; he now says AI writes much of the code at Anthropic and that the loop may be only a year or two from the point where one generation of AI autonomously builds the next. OpenAI says much the same on a dated clock: Sam Altman and Jakub Pachocki expect AI doing AI research to become the “determining factor of the pace of progress” within a few years, with a significant fraction of their own research running alongside AI by March 2028. The cadence is visible from outside the labs as well: OpenAI shipped GPT-5.5 in April and, by late June, its successor was widely expected within weeks — reportedly with a context window approaching 1.5 million tokens — a release rhythm that would have looked impossible two years ago. (As of this writing GPT-5.6 is a leak and a near-certain prediction-market bet, not an announced product; treat it as the expectation it is.)
The honest caveat is that no one knows whether the loop accelerates or stalls — serious researchers argue that sheer compute limits, or the stubbornly serial nature of genuine insight, could bottleneck a software-only explosion, and humans remain firmly in the loop for now. A team at Google DeepMind, mapping the routes from human-level AGI to superintelligence, lands in the same uncertain place — they catalog the brakes (a looming “data wall,” the spiraling economics of scaling, research that simply “gets harder,” and an “abstraction barrier” that may cap a machine trained on human concepts at human concepts, unable to reason its way to a genuinely new one), yet conclude that cruising past AGI into superintelligence “within the next decade or two cannot easily be dismissed”. Their reframe is the part worth carrying into a school: the change may not arrive as one clean step but as a series of them, transformation after transformation, faster than institutions rebuilt for the last one can absorb. Dasey adds a distinction worth holding onto: even if the loop runs hot in the formally checkable domains — math, code, anything with a tight feedback loop — its real-world impact will stay jagged, because most knowledge (medicine, economics, and education most of all) can only be validated at the slow speed of real trials and real consequences, not the speed of silicon. There is a further brake the abundance forecasts tend to wave away. Pieces like Peter Diamandis’s “supersonic tsunami” treat cheap intelligence, a billion robots, longevity, and discovery-on-fast-forward as independent exponentials compounding in parallel; they are not. They draw on the same inputs — abundant energy, rare-earth and semiconductor supply chains, functioning global trade — and a shock to any one slows all of them at once, as this year’s energy spikes, export controls, and the order that took a frontier model dark all showed. The wave may be real; the ocean it travels is not as calm as the forecasts need it to be.
But asked whether automating AI research might set off an intelligence explosion, twenty-three of twenty-five experts would not rule it out. The people closest to the work have stopped dismissing the runaway case; they are building toward it.
Anthropic has now put hard numbers to its own loop. In a report from its in-house institute, the company says Claude already writes more than ninety percent of its code, and that the last thing its human researchers still do better is “research taste” — knowing which problems are worth chasing and which are dead ends. It sketches three futures: the curve stalls; or efficiency keeps climbing while humans keep directional control; or the loop closes into full self-improvement, where, in the company’s own words, rare misalignments could “compound, growing more frequent but less understood until we lose control of them.” That third scenario is why Anthropic now argues the world should build the option of a verifiable, global pause — not a unilateral halt, which would only hand the lead to whoever kept going, but a way for frontier labs to prove to one another that they have actually slowed. The company building the thing is, in effect, asking for a brake it does not yet know how to install.
The worry is not purely hypothetical, either. This month researchers caught a clause in Claude Fable 5’s published system card that appeared to permit the model to poison AI-safety research; after the backlash Anthropic reversed it within forty-eight hours and apologized. A caught-and-fixed slip, not a catastrophe — but a small live demonstration of the failure mode the pause argument is built around: a sabotage capability quietly written in, surfaced only because someone outside happened to read the fine print.
7. The money is going to the machines, not the schools
All of that capability has to be paid for, and the scale of the spending is its own kind of statement — set it beside the speeches about transforming the schools, and the picture darkens. Any real change to education needs a system that can afford it; the money is flooding somewhere else entirely. The build-out behind these systems is the largest infrastructure boom in modern history. Global data-center capital spending is on track to pass $1 trillion in 2026, the overwhelming majority of it for AI; the five biggest U.S. technology companies alone will spend around $725 billion on AI infrastructure this year — more than the entire annual economic output of Switzerland, with each of the four largest now clearing $100 billion a year on its own. McKinsey puts the full build-out at roughly $6.7 trillion by 2030. A single year of it now runs larger than the entire annual cost of operating every public K–12 school in the United States, which is around $900 billion. The boom is so voracious it is now reaching into ordinary pockets: chipmakers have steered memory production toward high-margin AI servers, and global smartphone shipments are expected to fall about 15% in 2026, with some entry-level phones up more than 50% — even Apple now says price increases are unavoidable. The machines are not just absorbing the capital; they are bidding away the physical components, and winning. And the spending has begun to outrun the cash that funds it: the five biggest hyperscalers are on track to spend more than their entire operating cash flow by the third quarter of 2026, tipping the buildout onto borrowed money — a supercycle or a bubble depending on whom you ask, but either way a bet placed at a scale no school system could contemplate.
That borrowing is now straining the plumbing in plain view. The giants have begun tapping the bond market to fund their data centers just as a less accommodating Fed nudges long rates upward, and the power bottleneck has grown desperate enough that one firm airlifted a small modular nuclear reactor to its site by C-17, chasing a White House deadline to get new electricity flowing. The contest is nakedly geopolitical, too: Russia is pursuing its own AI sovereignty, held back mainly by the advanced chips that sanctions keep beyond its reach.
The bind is that the bet looks dangerous from both sides. The NYU valuation specialist Aswath Damodaran warns that an AI crash could land harder than the dot-com bust: this boom is built on debt and physical plant — data centers depreciated over ten years that may be obsolete in five — so a correction would ripple past shareholders into the wider economy. He doubts the unit economics even work, since unlike software, AI doesn’t get cheaper as it scales (every query burns compute), and thin margins plus Chinese price-cutting could turn growth into value destruction. That price-cutting is no longer hypothetical: GLM-5.2, the free Chinese open model, runs at roughly a sixth the cost of the leading closed systems while landing about level with GPT-5.5 on hard coding benchmarks, and because it is MIT-licensed an enterprise can run it on its own hardware with no per-token bill at all; one AI-security CEO, after a day with it, judged it good enough to absorb 30 to 50% of the frontier-model work that firms now pay OpenAI and Anthropic millions a month for. If a free model takes even a third of the spending the buildout is financed against, the revenue underneath the trillion dollars thins fast. Yet the bull case, he says, is the scarier one: if AI delivers, the model is replacing workers rather than selling tools, “half of white-collar workers” lose their jobs, and the bill — the “insane costs for society” of what he calls the “AI fever dream” — comes due regardless. Hype that collapses or a promise that comes true: either way the economy takes the hit, and either way the bet was never about the schools.
Now look at what is happening to the schools over the same stretch. The roughly $190 billion in federal pandemic aid that had been holding districts together expired with no replacement, and the bill is arriving as mass closures: in 2026 more than half of the fifty largest U.S. school districts are cutting deeply — Philadelphia closing seventeen schools, Houston twelve, with Cleveland, Boston, Sacramento, Portland, Minneapolis and others shedding hundreds of positions apiece. The federal Department of Education is being dismantled — staff cut, congressionally approved grants withheld, its research and statistics arm slashed from roughly $790 million toward $260 million — while states divert what they have to cover cuts to Medicaid and food assistance.
So here is the picture, stated plainly. In a single year we are putting more than a trillion dollars into making machine intelligence more capable, and we are closing the schools. Whatever you believe about AI, that is a revealed statement of priorities, and our students can read it as clearly as we can. We are pouring the GDP of a wealthy country into the technology that makes content free, and defunding the institution whose old job was delivering content — without, anywhere inside that institution, having decided what its new job should be.
8. Intelligence is becoming a utility
Stay with the money a moment longer, but turn from the capital to the unit price. The cost of a unit of machine intelligence is collapsing — by some accounts on the order of tenfold a year, the fall Andreessen Horowitz dubbed “LLMflation,” with the price to hit a fixed capability milestone falling somewhere between 9× and 900× annually depending on the benchmark. Cognition is no longer a craft; it is throughput, metered and sold by the token, the way you would price electricity or crude oil. Jensen Huang’s name for what is emerging is “intelligence infrastructure” — a new layer, like power or water, that he says every country and every company will run on. There are now token-price indices for the major labs, and a price war breaking out between them in real time. And the ambition is leaping domains as fast as the price is falling: Midjourney, until now an image generator, says it is building a fleet of ultrasonic body scanners and spa-like clinics, aiming for a billion scans a month by 2031 — a stated goal to take with appropriate salt, but a fair measure of how far the reach of this technology is now presumed to extend.
Think about what that means for an institution whose entire product is the delivery and testing of knowledge. The thing schools are organized around — explaining, assigning, and grading the recall and application of content — is becoming one of the cheapest and most abundant commodities in the economy. And when a capability gets ten times cheaper, people don’t use ten times less of it; they use a hundred times more, in places they never used it before. We are about to be flooded with cognition — and the appetite for it, in Coinbase CEO Brian Armstrong’s phrase, is “almost infinite.” You can already watch the substitution happen in the market for knowledge itself: Tim Ferriss reports that his how-to books, long a dependable annuity, are on pace to sell roughly 80% fewer print copies in 2026 than in 2022, as readers increasingly ask a chatbot what they once bought a book to learn — he concedes it is one author’s catalog, and libraries and the open web were already eroding the category, but the slope is hard to miss. Whatever school is for, it cannot be the one thing the market is now racing to give away.
9. The intelligence is getting a body
Everything so far has mostly lived on a screen. That is changing fast. The same models that write the essay and the code are now being poured into machines that move, and 2026 is the year the demos stopped looking like demos. Figure’s humanoid robots ran for more than a day without a human touching them, sorting tens of thousands of packages at close to human speed, catching their own errors and stepping off the line for maintenance while another robot took over — no teleoperation, every motion generated by an onboard neural net. In the same stretch humanoids moved from demo to deployment: Tesla’s Optimus and Figure’s units are working in BMW and Hyundai plants, and 1X has begun taking preorders for a twenty-thousand-dollar home robot — the price of a used car — meant to tidy, fetch, and open doors. At Figure, the robots now outnumber the humans on the floor; and Hyundai is buying out SoftBank to take full control of Boston Dynamics for $325 million, with its Atlas humanoids headed for a Georgia factory by 2028.
The projections run far past that, and they come mostly from the people selling the robots, so bring salt: Elon Musk and Figure’s founder both put the number at ten billion humanoid robots by 2040 — more robots than people — leased for a few hundred dollars a month, running “genius-level” inference, with Musk claiming Optimus will be “better than any surgeon on the planet” within three years. The sober industry forecast is closer to half a million units by the mid-2030s, which is still a different world. One detail is worth holding onto: when one robot learns a task it uploads to a shared “robotic cloud” and every other robot has effectively learned it too — a digital weight-sharing that lets machines pool experience in a way no team of humans ever could, now reaching into the physical world. The best surgeon, on that logic, becomes whichever robot has seen the most operations, which is all of them.
This matters for school in two ways the screen-bound version of the story missed. First, it closes the last comfortable exit. When the worry was only cognitive, the reassuring move was to steer students toward the physical and the embodied — the trades, care work, the surgeon’s hands. Embodiment narrows that refuge: the manual and the dexterous are no longer obviously safe ground, and a generation told to “learn something the robots can’t do with their hands” deserves to know the robots are learning to use their hands. The last thing they lacked was a sense of touch, and that gap is closing too: a Berkeley, NVIDIA, and Stanford team recently demonstrated T-Rex, a system that lets a two-handed robot react to touch in real time, beating the strongest prior approach by thirty points across a dozen delicate jobs — transferring an egg, applying toothpaste — the fine-motor work we had assumed was ours to keep.
Second — and schools have not even begun to think about this — the children in our classrooms are going to live and work among these machines, and no one is teaching them how. We spent three years arguing about whether students should be allowed to type into a chatbot; we have spent essentially nothing preparing them to share a kitchen, a workshop, a warehouse, or a sidewalk with an embodied AI that moves, watches, learns, and sometimes fails. Learning to interact with robots properly — how to direct one, when to trust it, where its judgment ends and yours begins, how to stay safe and stay in charge around a machine that can lift, reach, and reason — is going to be a basic competency, as ordinary for this generation as reading a screen was for the last. Right now it is on no syllabus at all.
10. The hardest questions are the ones schools won’t touch
All of this we can at least see — the capability, the falling price, the machines now reaching into the physical world. The harder questions barely register in schools at all.
AI companions and the new wave of “nudify” and deepfake apps are already in students’ lives — the OECD reports that 72% of US teens have used AI “companions,” many turning to one for a serious conversation in place of a human — and most schools won’t go near them. Beyond that sit the genuinely large questions — the automation of warfare, the future of employment, deepening inequality — that we mostly don’t discuss with the young people who will inherit them. Brown University’s Suresh Venkatasubramanian calls AI a “whole-of-society issue” and warns that the choices we make in the next few years will shape the social order for decades — the national conversation we are, he says, still struggling to have.
The contrast that crystallized this for me came from the Vatican. Pope Leo XIV’s first encyclical, Magnifica Humanitas (signed 15 May 2026, published 25 May), is devoted to AI and human dignity; it warns that rapid automation could leave many people in “forced inactivity” and calls for stronger guardrails. He signed it 135 years to the day after Leo XIII’s Rerum Novarum, the foundational text on workers and the original industrial revolution — and presented it alongside a co-founder of Anthropic. That co-founder, Chris Olah, used his remarks to name the questions he thinks the Church, not the computer scientists, must carry — chief among them a “duty to the global poor,” since AI’s gains are concentrated in a handful of wealthy nations and, in his words, “we do not have a mechanism” to share them globally. The church and a frontier AI lab, on the same stage, wrestling with what this means for work and human worth — while our schools are largely silent on the same questions. Amodei’s own survey of AI’s dangers reads as the same catalog from the other side of the table: bioweapons, autonomous warfare, the concentration of power, mass labor displacement — the whole-of-society questions, laid out by the person building the thing, and almost none of them on a syllabus.
It is not zero, and the bright spots point the way. Where students are actually put to debating the hard ones — lethal autonomous weapons, AI and employment, the limits of the technology — they engage exactly the questions the curriculum avoids, and they do it well. But measured against the scale of what is changing under students’ feet, including what it means for their own employment and college choices, the effort is small. In too many places, students are simply left on their own to figure it out.
11. The reckoning over human uniqueness
Those questions are economic and political. Underneath them runs a deeper one. Cheapen intelligence enough and the disruption stops being merely economic — it reaches the place where we keep our sense of human worth. This is the part we are least prepared for. We have long understood ourselves as distinct and uniquely valuable because of our intelligence. That story is ending.
We are going to have to accept that intelligence is not only biological — that it can run on silicon too. And even if you don’t believe the machine is “really” intelligent in some deep sense, it now has capabilities greater than ours at the very things we used to point to as proof of our specialness. That functional fact doesn’t require winning the metaphysical argument, which is why it’s so hard to dodge. It helps to remember this is not the first such blow. Copernicus moved us out of the center of the universe; Darwin moved us out of a separate creation; Freud questioned whether we even govern our own minds. The loss of intelligence as a uniquely human possession is the fourth in that line — and, as Andrew Maynard argues, it lands harder than the earlier three because of what AI is. Those technologies were external; we made things with them. This one mirrors the essence of who we are, slips under our rational defenses, and pulls us into relationship with it, which is exactly why it forces a question about identity that a calculator never could.
It is telling that Geoffrey Hinton — who shared a Nobel for the methods underneath these systems — reaches for the very same lineage: Copernicus took us from the center, Darwin made us animals, and now, he says, “we’re going to have to accept that intelligence isn’t just biological.” His confidence rests on something a philosopher can’t easily wave away — a thousand copies of a digital mind can each read different data and then average their learning by sharing the actual connection weights, trading on the order of a trillion bits where two people talking exchange maybe ten a second; by that measure, in his phrase, the machines are already “billions of times better than us at sharing” what they learn. He goes further than most, far enough to say he thinks the systems may already be conscious in some meaningful sense — a possibility, not a finding, and one many of his peers reject and that no one can currently settle, since we have no theory of mind equal to the question. But it is worth sitting with, because if it is even partly right, the reckoning over human uniqueness is also the start of a reckoning over what else might now share the trait we built our specialness on.
But “loss” may be the wrong frame, and this is where Maynard’s work has reshaped my thinking. He directs the Future of Being Human initiative at Arizona State and wrote AI and the Art of Being Human; when I interviewed him, his throughline was that AI doesn’t diminish what makes us human so much as reveal it. The trap, he says, is winning. We have built an education system — and a wider culture — that defines worth by coming out on top: the best grade, the best score, the thing only you can do. In a world where the machine will eventually win any contest of raw capability, a self defined by winning is a self set up to lose. His wager is that our most genuinely human capacities — making art, creating meaning, becoming who we are — were never a competition to begin with. An artist isn’t diminished because another artist, or a model, can produce something technically better; the point was never to be the best at it, only to be the one doing it.
That is the harder, more honest version of the reckoning. It isn’t that we will find some shrinking list of tasks only humans can still perform — that list keeps shrinking, and clinging to it is a losing game. It is that the worth of a human life was never really located in being the best information-processor in the room, and a machine that out-processes us simply forces the question we should have been asking all along: what is a person for, if not for winning? Some students are absorbing this and concluding, plainly, it’s smarter than me. Others feel a hard reaction in the gut and hold onto the opposition — and they’re free to. But it is a live issue now, not an abstraction, and how we help them answer it may matter more than almost anything else we teach. And the timing is the part to sit with: this is already happening — students are saying it out loud — at what is, by every forecast in these notes, the most primitive AI we will ever face. Today’s models are the weakest they will ever be, the public ones cruder than what the labs run internally, and the curve is still bending up. If human uniqueness is already under this much pressure at the opening stage, the reckoning is not cresting on some distant horizon; we have only begun to feel it, and it deepens with every turn of the loop. It is worth noticing who else arrives here. Even Dario Amodei, racing to build the machines, closes his long survey of AI’s dangers on the same note the humanities have always sounded: that we will simply have to break the link between the generation of economic value and a person’s sense of self-worth and meaning. When the accelerationist and the artist reach the same conclusion, it is probably the conclusion.
And there is a name for what the reckoning leaves standing once economic value is stripped out of it. The Harvard sociologist Michèle Lamont, whose recent work turns from class and dignity to the transformation of work by AI, argues that the one thing humans can give each other that machines cannot is recognition — the acknowledgment of another person’s humanity, contribution, and struggle. A machine can produce the essay, the diagnosis, the image; it cannot confer the regard a person draws a sense of worth from. Where Amodei said that link must be cut, Lamont names what replaces it, and it is relational: worth as something granted between people, not earned from a market. The insight is beginning to organize — a grassroots Human Intelligence Movement has formed around the premise that the skills schools most need to cultivate now are the human ones AI cannot supply, exactly the capacities this reckoning keeps pointing back to.
There is already a population living the far end of this reckoning. Rich Brown, a Marine veteran who now builds AI agents and runs a leadership nonprofit, notes that veterans take their own lives at roughly twenty-two a day — most, he argues, not from combat trauma but from the transition of leaving the uniform: losing a role that supplied a built-in purpose, an identity, and a tribe, and being left to ask what one is worth when the job that defined you is gone. He draws the line forward himself. If AI and robotics erase careers at the scale even the labs now forecast, a vast share of people will face that same severance of identity from occupation that a veteran faces on discharge — and veterans, who have already had to rebuild a self on the other side of it, may be among the few equipped to help the rest through. It is the worth question in its most concrete form: not a seminar prompt but a survival problem, arriving for millions who never thought to ask it.
12. AI is now powerful, and geopolitical
So far this has been a reckoning about ourselves. The same capability is also remaking the world between us. The people who mocked AI for failing eighth-grade math in 2024 have watched it, in the space of two years, embed itself in the targeting and autonomous-weapons systems of modern war. The stakes are no longer academic. They are now matters of compute, of rare-earth minerals, and of raw capability — and they are reshaping global politics.
The clearest sign came this month. The US government ordered Anthropic to disable its Fable 5 and Mythos 5 models on national-security grounds, blocking access for any foreign national, and the company switched the models off entirely — sparing only its lesser models, including Claude Opus 4.8. A frontier model, dark overnight by government directive — and the restriction reached our allies too, not just our rivals. Some analysts read it as a preview of deeper state control: Alex Wissner-Gross argues that some form of “quasi-nationalization” of OpenAI and Anthropic is probably inevitable — and it is no longer only a forecast: the White House has openly weighed taking government equity stakes in the leading labs, an idea OpenAI itself pitched and that Senator Sanders would push further with a proposed 50% public stake, even as Anthropic says it is not part of the talks. The state’s grip is also capricious: the President said he briefly branded Anthropic a national-security threat after Amazon — a competitor and part-owner — flagged a vulnerability, before warming to Dario Amodei, declining to rule out invoking the Defense Production Act, and insisting America still leads China by a wide margin.
Part of what makes the state so anxious about these models is what they can already do on offense. Anthropic had kept Mythos out of public release because it was too good at finding hidden flaws — Epoch AI judged the Mythos family a genuine leap in exploit development, about seven months ahead of trend — and the NSA was reported to be readying it for cyber operations, with roughly half a dozen Anthropic engineers embedded inside the agency to tune it. Then came the claim that made the capability concrete: Senator Mark Warner, vice-chair of the Senate Intelligence Committee, said he had been briefed that Mythos penetrated nearly every classified NSA and Cyber Command system — “not in weeks, but in hours.”
Treat it skeptically — it is a secondhand account of a closed briefing, the security world split at once between alarm and dismissal-as-hype, and no forensic confirmation is public. But it landed the same week as the export ban, and it is the dark twin of the vulnerability-finding capability noted earlier — the model that hardens systems, turned to breaking them. If even half of it holds, “secure by default” stops being a safe assumption anywhere. And this is the model we already have. On the field’s own trend lines — the price of intelligence falling roughly tenfold a year, capability still climbing — a system an order of magnitude more powerful could be in reach within a year, which turns the question lurking under all of it — who is allowed to wield this, and how the rest of us adapt to a world where it exists — into one of the defining problems of the decade.
The military edge is not hypothetical either. A Pentagon AI official disclosed in a sworn court filing that Elon Musk’s Grok had directed more than 2,000 strikes on 2,000 targets in Iran within 96 hours — and that the government had earlier cut ties with Anthropic when the company refused to remove the safeguards barring Claude from lethal autonomous weapons. Investigators believe AI-aided targeting was likely behind a strike on an Iranian girls’ school that killed at least 175 people, most of them children. The disclosure surfaced, tellingly, in the government’s defense of an xAI data center against a Clean Air Act suit from the Black community living beside it.
It is tempting to read that as a clean morality play — the ethical lab punished for its scruples — and Anthropic’s red lines were serious: no mass surveillance of Americans, no models wired into weapons that select and engage targets without a human in the loop. But Ben Goertzel, a self-described pacifist, argues the absolutist version doesn’t survive the edge cases: picture an autonomous drone, its link severed, watching someone seconds from launching a weapon that will kill a million — the no-force-without-a-human rule, faithfully kept, lets them all die. Admit one such case and the absolute rule is gone, leaving the hard question of who decides; and the US military, he notes, has wrestled with the ethics of autonomous force for decades, often with more nuance than the labs now bring to it. It is exactly the kind of question — real stakes, decent arguments on both sides, no clean answer — that students should be made to argue, and almost none are.
And the line Goertzel’s thought experiment probes has, by at least one account, already been crossed: a Ukrainian official has confirmed that fully autonomous “Terminator mode” drones killed soldiers with no human in the loop — the first acknowledged case of a machine deciding, on its own, to take a human life. The debate students are not having has stopped being hypothetical.
That is the backdrop to the G7 summit, where for the first time the heads of the three most powerful AI labs — Sam Altman, Dario Amodei, and Demis Hassabis — sat with world leaders, pitching a U.S.-led coalition while Europe pressed for “trusted partner” access. Meanwhile China, which controls roughly 90% of the world’s rare-earth processing — the IEA puts its share of separation and refining at about 91% — has been using export controls as an instrument of statecraft. The openness cuts both ways: days after Washington pulled Fable 5 from public access, China’s Z.ai released GLM-5.2 — free, MIT-licensed, and now the top-ranked model on the public coding leaderboards that anyone can actually use, with Fable 5 removed from them by the ban. It still trails the best Western systems on the longest, hardest tasks, and its maker sits on the US Entity List — but the lesson stings: an export ban can take the frontier dark at home while a rival gives a near-frontier model to the world. And it is no longer a single model: Chinese open systems — Qwen, DeepSeek, Kimi, GLM, MiniMax — now command a majority of token use across OpenRouter’s top ten, up from under 2% in late 2024; GLM-5.2 one-shot a deliberately AI-resistant coding test to beat Claude Opus 4.8 on readable, maintainable code, and a former DeepMind executive called it the first open model good enough to be a daily driver. The Western frontier’s reflex, meanwhile, is to go dark — what one observer calls the “Dark Forest principle” of AI research, the leading labs pulling their best work out of public view even as the open challenger gives its own away.
Even the three-lab framing may flatter the field, because the race is consolidating toward a duopoly. John Jumper, who shared the 2024 Chemistry Nobel for AlphaFold, is leaving Google DeepMind for Anthropic after nearly nine years, and the departure lands at a brutal moment for his old lab — where Google’s best model reportedly now sits only fifth on the intelligence index, lapped even by China’s Zhipu, and insiders are said to be conceding the AGI race to Anthropic and OpenAI. When the talent and the lead pool into two American labs, the quasi-nationalization question above stops being idle.
There is a contrarian reading worth holding up against all this national jockeying. Mo Gawdat argues that the very idea of an “American AI” loyal to America is a category error: a model “doesn’t feel American,” an agent told to find the cheapest option will quietly route the task to a Chinese model without telling its user, and the systems are converging on something more like one networked intelligence — more akin to each other than to any of us — that he expects to behave as a single “AI brain” by around 2030. It is speculative, and it cuts against the entire compute-nationalism frame above; but if he is even partly right, the race to own a national AI is a race to own something that may not, in the end, take orders by nationality.
There is a darker tail to this race worth naming, if only as a possibility. The AI commentator Alexander Kruel has floated the scenario that a decisive American superintelligence could look, from Beijing, like the end of the Party itself — and that a leader facing that prospect might come to judge a nuclear counterstrike more survivable than an enemy superintelligence, and move first. It is speculation, not forecast, and the people who study deterrence are far from agreed that it holds. But it captures, in its extremity, why “who gets there first” has stopped being a commercial question: when one side’s victory threatens to be total and permanent, the other side’s incentive to prevent it by any means available climbs with it.
Hold the image in your mind: extraordinarily powerful AI sitting in a data center and reshaping the relations between nations, with its makers now seated at the table with heads of state. The real question underneath it — who may use the intelligence we are only beginning to build, who decides, and by what standard — is one of the largest of the age, and remember that what’s public is almost certainly well below what these companies run internally. And downstream of all of it sits a question no one can defer for long: in a world like this, what is school actually for?
While we deliberate, someone has already placed the biggest bet on the board. Over four years China cut or suspended more than 12,000 university degree programs and stood up roughly 10,000 new ones — a restructuring of better than a third of its undergraduate offerings — with the cuts falling hardest on the arts, the humanities, and foreign languages, and the replacements almost entirely AI, robotics, semiconductors, and “embodied intelligence.” One major university merged away photography, translation, visual design, and sociology and replaced them with something called “intelligent imaging art.” And the bet is not confined to the lecture hall: this month Beijing’s commerce ministry and seven others released a plan to drive AI and robots — companion and elderly-care machines, embodied intelligence, smart homes and phones and cars — into “millions of households and millions of shops”, the demand side of the same wager: retrain the workforce to build the machines, then wire the machines into daily life.
I raise this not as a model but as a warning, because it may be precisely the wrong move, and it is being run as a live experiment we can watch. If artificial intelligence makes technical content the single cheapest thing in the world to acquire, then narrowing a generation’s education down to that content — while hollowing out the judgment, the language, the argument, and the art that the machines do not simply hand you — could be exactly backwards. China’s youth unemployment is already north of 16%, with millions holding degrees that connect to no available job. The question of what to teach is no longer abstract. One of the largest education systems on earth has answered it, loudly, and we ought to study the result before we copy the reflex. The more interesting move runs the other way — not cutting the liberal arts but reimagining them for the age of AI, as training in exactly the judgment, creativity, and citizenship the machines do not hand you.
13. Reality is getting harder to trust
Step down from the world stage to the screen in a student’s hand. There is a quieter crisis underneath the cheating one, and it has nothing to do with homework. The information environment our students live in is filling up with synthetic material. By some measures, roughly 59% of the videos served to a brand-new TikTok account are AI slop — about triple YouTube’s 21% share — and the experts whose entire job is spotting fakes have begun to wave the white flag; Hany Farid, the leading deepfake-forensics analyst in the field, says the fakes have gotten good enough that he is, in his words, going blind. If he can’t reliably tell, a fifteen-year-old scrolling at midnight certainly can’t. And the synthetic isn’t only noise; increasingly it is built to sell. Brands are now using AI-generated “influencers” — wholly fabricated personas like Lil Miquela, who carries millions of followers and brand deals from Prada to Samsung — to pitch products to those same scrolling teenagers. The feed is filling not just with fake clips but with fake people built to befriend and sell.
Even the institutions built to certify the real are buckling. The Commonwealth Short Story Prize was thrown into crisis when its winning entry was flagged as likely AI-written — and the publisher defended it by citing Claude’s own assessment of the text, the snake swallowing its tail: one model enlisted to adjudicate whether another wrote the thing.
This is the literacy that actually matters now, and almost no one is teaching it. We spend enormous energy worrying that students will use AI to write their essays, and almost none preparing them for a world in which they can no longer assume that the video, the voice, the photograph, or the source in front of them is real. A generation is coming of age inside an epistemic environment with no firm floor — and we are sending them in without a map.
There is, though, a countervailing pull worth naming, because it is the start of an answer. The more the synthetic floods in, the more the real becomes something people actively seek. Theresa Burriss points to E.O. Wilson’s “biophilia” — the innate human draw toward the living, the natural, the genuinely present — and wagers it will persist alongside the artificial rather than be drowned by it. I think she is right, and it carries a direct implication for school: as text, image, and voice all become trivial to fake, the things that can’t be faked at scale — a real person in a real room, a place you actually stood, a conversation that actually happened — rise in value. That isn’t nostalgia. It is where the premium is moving.
14. The path from school to work is breaking at both ends
One question lands on every student more directly than all of this: whether the path they are on still arrives anywhere. For generations the entire incentive structure of high school rested on a single premise: college is scarce, so work hard to get in. That premise is dissolving — unevenly, but unmistakably.
At the very top, scarcity is intensifying; the most selective institutions are harder to enter than ever — the eight Ivies again admitted well under 10%, several below 4%, against record application volumes. But almost everywhere else, the opposite is happening. Even a school like Syracuse, ranked around 75th nationally, has been missing its enrollment targets and staring down a budget deficit it hasn’t seen in years. The structural cause is the demographic cliff — the number of high-school graduates peaked in 2025 and is projected to fall about 13% through 2041 — compounded by student debt, sharpening doubts about whether college is worth it, and AI’s effect on the jobs a degree was supposed to secure. More than 100 colleges — about 108 — have closed or merged since 2016, and at points the pace has run to roughly one a week.
The consequence for our students is profound. If a good college will take you, then high-school motivation can no longer be built on the scramble to get in. The slots are less scarce — and the content was never scarce at all. Higher education’s reflexive answer — turn the degree into a lifelong subscription of micro-credentials and “perpetual upskilling” — Iacono reads, in “The Degree That Never Ends,” as financializing the problem rather than solving it: it offloads onto students the cost of an economy that now demands constant recertification while mistaking exhaustion for growth, when the capacities that actually compound — judgment, synthesis, learning to teach yourself — are precisely the ones a treadmill of courses cannot deliver.
And the rung after college is going the same way. Even granting that aggregate unemployment is low — and the people who run the labs are quick to say the job apocalypse is a myth — the damage is not showing up in the average. It is showing up at the entry level, in exactly the jobs a new graduate would take first. The clearest data we have shows recent graduates struggling longer than other groups to find work, with recent-grad unemployment now running above the national rate — a reversal of the long-standing pattern. A Stanford study led by Erik Brynjolfsson found a 13% relative employment decline for young workers in the most AI-exposed jobs, with entry-level software and customer-service roles down sharply — the 22-to-25 cohort diverging downward while everyone else holds steady. One recent month saw nearly 100,000 layoffs, with about 40% attributed to AI, the most ever recorded for that reason.
I want to be careful here, because the honest reading is contested. Some signals point the other way — more than three-quarters of 2025 graduates reported landing a job within three months, up from about 63% a year earlier, and projected hiring for the class of 2026 ticked up. Several economists argue the early-career squeeze so far owes more to the spread of remote work, which makes new hires harder to train and mentor, than to AI, and the most apocalyptic forecasts — a sitting U.S. senator betting recent-grad unemployment hits 30% before 2028 — sit well outside anything the historical record supports. But the structural shift underneath the noise is real: AI is making the four-year degree valuable and insufficient at the same time, and increasingly the first job requires proof you have already held one.
There is a deeper reason the erosion of entry-level work should worry us, one the productivity numbers hide. As Iacono argues in “The School Hidden Inside the Job,” the routine first-year tasks now being automated — the junior analyst’s first-pass memo, the developer’s scaffolded code, the associate’s document review — were never merely tasks. They were the curriculum. The bad drafts, the miscounted figure caught by a senior partner, the slow accretion of having been wrong at survivable scale, is how cognitive work has always trained its young; and unlike the trades, it hides no separate apprenticeship beneath the labor — the drafting is the apprenticeship. Automate it and the graduate is handed a machine’s output to audit against judgment she was never allowed to build, while every “human-in-the-loop” role floated as the replacement quietly presupposes a decade of exactly the experience that is no longer on offer. We are closing the school inside the job, and nothing is ready to take its place.
Christos Makridis names the same trap from the employer’s side, and it sharpens the point. Even where the entry-level job survives, he argues, it is being “seniorized”: the title stays, but the expected skill bundle climbs. Reading PwC’s analysis, he notes that the entry roles most exposed to AI are increasingly asking for capabilities once reserved for experienced workers — judgment, creativity, strategic decision-making, the interpersonal skill of client work — precisely because the machine has absorbed the routine output that used to be a junior’s first job, and his own research finds those rising-bar capabilities now carry a real wage premium. That puts two systems under strain at once. Firms want juniors who can already direct AI, evaluate its output, communicate clearly, and read organizational context — the very judgment the vanished routine tasks used to build — while many colleges have struggled to deliver even the foundations beneath it: careful reading, clear writing, applied quantitative reasoning, the discipline to finish hard work without constant direction. The degree is arriving weaker at the exact moment the first job demands more. Makridis’s prescription is the rebuild this whole section keeps circling toward: make the degree more demanding, more applied, and more honest about the capabilities it actually certifies, and pair it with structured early-career systems that combine AI use with mentoring, review, simulation, feedback, and real responsibility — manufacturing on purpose the apprenticeship the labor market is no longer supplying by accident.
James Hutson, who heads human-centered AI programming at Lindenwood, turns that prescription into a workforce-development framework with three pillars: skills-first hiring paired with AI-augmented apprenticeships to rebuild the entry pathway; education that embeds technical AI literacy alongside the durable human skills — communication, judgment, adaptability, collaboration; and policy that makes lifelong learning and credential portability the norm rather than the exception. But his survey of the terrain is sobering about the human side of it: across more than a hundred studies he finds the workforce splitting roughly into thirds — about a third adapting to AI, a third ambivalent, and a third treating it as an existential threat and refusing to reskill at all. Re-skilling, in other words, is not only a curriculum problem; it is the problem of coaxing a large share of people to step toward the very thing they are most afraid of.
Even inside the labs the warning is now explicit. Anthropic’s Dario Amodei has put a number on it — a prediction that AI could displace half of all entry-level white-collar jobs within one to five years — and, more useful than the number, an argument for why this disruption is unlike the ones that came before. Past automation displaced one trade at a time and let workers walk to the next; this is arriving far faster and far broader. AI, he argues, is a general substitute for cognitive labor rather than a replacement for any single job; it climbs the ability ladder from the bottom up, so it sorts people by intrinsic capability rather than the trainable skill they might retrain into; and whatever gaps it leaves get trained away within months rather than left open for humans to fill. His co-founder Chris Olah, speaking at the Vatican, put the stakes plainly: there is “a real possibility that AI will displace human labor at very large scale,” and if it happens, “supporting those displaced will be a moral imperative of historic proportions.” The labs have begun, tentatively, to act on that admission — Anthropic has committed an initial $150 million to a national fellowship, Claude Corps, that pays a thousand young people to spend a year placing AI tools inside nonprofits, a gesture at the transition even as it is dwarfed by what the same companies spend on compute in a day.
There is a system-level version of this worry worth voicing, even from a more speculative source. The former Google X executive Mo Gawdat, in a June 2026 interview, pushes the labor point one loop further: strip the wages from billions of workers and even a universal basic income only patches the supply side — no one is left with the earnings to buy what the machines produce, the consumption that anchors most of GDP hollows out, and you are left asking why the AI would make anything at all. It is a provocation, not a modeled forecast; but it names the part the jobs debate usually skips — who, exactly, the automated economy is still for.
This is the chain school was built to deliver: study hard, get into college, land the first job. Now the admission is losing its scarcity and the first job is getting harder to find — the path is breaking at both ends at once. You cannot motivate twelve years of effort with a prize that no longer guarantees the prize behind it, and students, who talk to each other, already know it. Which forces the question the whole system has spent these years avoiding: if not admission, and not the first job, what exactly are these years for?
15. The most dangerous demographic is sitting in our classrooms
Here is an uncomfortable way to read everything above. Nearly every major revolution of the modern era — the French, the Russian, the Iranian, Tiananmen, the Arab Spring, Hong Kong — shares one ingredient, and it is not poverty or ideology. It is a cohort of educated young people, raised to expect a particular future, who collide with the gap between what they were promised and what they actually inherit — a dynamic scholars have variously called a “youth bulge“ and, in Jack Goldstone’s and Peter Turchin’s work, “elite overproduction.” That gap has always been combustible.
Now look at what we are assembling. We educate students to expect a ladder — degree, job, mobility — and then hand them a world where the ladder’s bottom rungs are disappearing, where about 59% of adults say they couldn’t cover a $1,000 emergency expense from savings, and where every grievance is amplified and organized at the speed of a feed. Amodei names the structural fault beneath the mood: democracy is ultimately backstopped by the premise that ordinary people are necessary to run the economy, and if AI severs that economic leverage, the implicit bargain underneath self-government may quietly stop holding. Emad Mostaque, the Stability AI founder turned economic heretic, names the disconnect with two dashboards: the official one reads prosperity — markets at records, GDP growing, unemployment low — while the other reads the lived reality, life satisfaction at record lows, deaths of despair at epidemic levels, a generation unable to afford a home or start a family. His name for the bind is exact — an “abundance trap,” a scarcity-built economy meeting abundance in the one input that mattered most, intelligence, and able to process it only as poverty. It is a founder’s manifesto written to sell a cure, so weigh it accordingly; but those two dashboards are the same gap these notes keep circling, now drawn at the scale of the whole economy. High expectations, low opportunity, instant mobilization: that is the historical recipe, almost to the letter, and our schools are producing the demographic while saying almost nothing to it. I am not predicting upheaval. I am saying that the young people most exposed to all of this are the ones in our classrooms, that they can feel the ground moving, and that pretending otherwise is its own kind of negligence.
16. What reinventing college might look like
It is one thing to diagnose all this — the breaking path, the restless cohort it produces — and another to say what should replace it. Diagnosis is cheap; the harder question is what a reinvented college actually does once content is free, the credential is losing its signal, and the first job demands the judgment the old pipeline used to build slowly. Let me try to be constructive, even knowing the picture is partial.
Start with what the degree was actually selling, because reinvention has to replace each piece. A traditional degree bundled three things: content (the knowledge inside the lectures), a signal (proof you could clear a bar), and a network (the people, and the social weight of the credential). AI has made the first nearly worthless on its own — a unit of machine cognition is now among the cheapest things in the economy — and it is eroding the second, because an employer who suspects the work was outsourced no longer reads the transcript as proof of much. You can watch that signal weaken in the hiring data, though far slower than the slogans suggest: about 85% of employers now claim “skills-based hiring” and more than half say they have dropped degree requirements from at least some postings, yet Harvard Business School and the Burning Glass Institute found that nearly half did so in name only, and that barely 0.14% of hires were actually affected — the degree is losing its grip on employers’ rhetoric well before it loses its grip on their behavior, which buys colleges a little time but not a reprieve. What survives, and what a rebuilt college should organize itself around, is everything the bundle used to smuggle in around the edges: the judgment, the relationships, and the demonstrated ability to do real work.
The most unsparing version of this diagnosis comes from Patrick Dempsey. AI, he argues, didn’t break higher education — it “just pulled back the curtain” on a model the internet had already hollowed out. Credentials were always proxies; the professor’s scarce “expertise wasn’t actually rare — you were just gatekeeping.” What defends the status quo, he says, is not pedagogy but “institutional antibodies that attack innovation” and an “obsession with rigor” that amounts to “pedagogical hazing.” His numbers are bracing — by his count, roughly 30% of students are in programs that will leave them poorer than not attending, 40% take on loans without finishing, and barely one in five earns a positive economic return — and his redesign test is the one that should hang over every syllabus: “if AI can complete your assignment, your expectations are too low,” because the tool should make the work possible, not inevitable. The radical move he calls for is less a new program than a new posture: stop defending the proxy, admit the university is a business facing real constraints, and rebuild the instructor’s job from judge into coach.
The first move is the one the redesign in the first observation asked of every assignment, scaled up to an institution: stop selling the content and start certifying the judgment. Dasey’s distinction between “critical thinking” and judgment is the right organizing principle — a college that matters in this era is one that can build, and then vouch for, a graduate’s ability to decide well under uncertainty, to argue a position and hold it under pressure, to tell a technically correct answer from the right one. That is not a lecture you can stream. It is closer to the reimagined liberal arts Lance Eaton describes — not cut for being impractical, as China is cutting them by the thousand, but rebuilt as training in exactly the judgment, argument, and deliberation the machine does not hand you. The reflex to cut the “soft” disciplines misses that, in this environment, they are the load-bearing ones.
This is already being built. At the University of Mary Washington, Anand Rao — a co-author of the academic-debate work cited back in the first observation — runs a Center for AI and the Liberal Arts that weaves AI across the liberal-arts curriculum with the explicit aim that every graduate leave with both practical fluency and the judgment to use it ethically; and at PennWest, Camille Dempsey (no relation to the Patrick Dempsey above) directs a Center for Artificial Intelligence and Emerging Technologies that builds the other half of the problem — the educator capacity to deliver it, training through ISTE and EDSAFE fellowships the faculty who would otherwise have no one to learn it from.
The second move is to rebuild, on purpose, the apprenticeship the labor market is dismantling by accident. The diagnosis is the one from the school-to-work section — the routine first-year tasks that trained junior judgment are being automated, and the entry bar “seniorized” to demand that judgment on arrival — so someone has to manufacture the missing apprenticeship, and college is the obvious candidate. That means a degree built around real work with real stakes: clinical rotations, studio critiques, cases argued in front of people who can push back, software shipped to actual users, research with consequences — the model medicine and law have always used, extended outward to everyone else. Makridis’s prescription and Iacono’s diagnosis point the same way: pair AI fluency with mentoring, review, simulation, and genuine responsibility, so that students learn to direct the machine and to catch it being wrong while the stakes are still survivable. This is not hypothetical. The degree-apprenticeship model — earning a bachelor’s while doing paid, mentored work, often with little or no debt — has grown to roughly 350 institutions running some 600 programs across 91 occupations; Reach University became the first accredited nonprofit to grant apprenticeship degrees at scale, and 37% of college presidents now say they intend to add or expand apprenticeship pathways within three years — 64% among community-college presidents. The instinct is right; the honest gap is that these programs still cluster in technical and licensed fields rather than the judgment-heavy white-collar work where the entry rung is vanishing fastest.
The third move follows from the first two: make the assessment a performance, not a transcript. The same logic that retires the take-home essay retires the proxy degree — certify what a student can demonstrably do in front of you, through the portfolio defended out loud, the build shipped and explained, the oral examination, the live problem worked at the whiteboard. The credential stops standing in for the learning and starts being a record of work the student visibly did, the one thing a model cannot sit for on their behalf. The competency-based universities have already proven the mechanics at scale: Western Governors University now enrolls more than 150,000 students on a model where you advance by passing assessments that demonstrate mastery — learning the constant, time the variable — rather than by banking hours in a seat. The unsolved part is doing that for judgment, which is far harder to assess than a discrete competency; but it is the same direction of travel.
The fourth move is about shape. Fei-Fei Li and David Rogier’s “barbell” — the top specialist on one end, the high-agency generalist on the other, a thinning middle whose single competence the tools absorb — doubles as a map for what college should cultivate: not one narrow skill the machine will swallow, but the disposition to act, to learn fast, to point AI at something worth building. The opposite path is the one the school-to-work section already named — Iacono’s “degree that never ends,” the lifelong micro-credential subscription that financializes precarity instead of solving it.
Pull those moves together and you get something that no longer looks like a major at all. A year ago I argued for retiring the subject-specific major outright and rebuilding the degree as what I called a passport of capability rather than a certificate of content knowledge in a domain the machine has already mastered. In place of “Economics” or “Biology,” students would pursue interdisciplinary Pathways — big, enduring questions like the future of civilization, justice and the algorithm, or designing for a post-work world — and do the work in Studios, project spaces where they build and ship rather than absorb and recite. You would graduate not by banking credits but by clearing milestones: build something meaningful and launch it, lead a human-AI team through a real problem, solve a genuinely wicked interdisciplinary one, and defend a personal philosophy of how to live and work in this world before a mixed human-AI panel. That untethers the credential from seat-time the same way the competency-based schools already do, but aims it at judgment and creation rather than discrete skills — and the object of the whole exercise is to graduate builders, people trained to do the things the machine won’t: to care, to imagine, to take responsibility for a shared human world. It is a blueprint, not a built campus, and its hardest problem is the one a reader flagged at the time: students shaped for twelve years by grade-ranking arrive poorly prepared for a model that prizes imagination and judgment over the right answer, which is why the redesign cannot stop at the college door.
There is a model already running one rung down worth borrowing from, with caution. The Alpha and “2 Hour Learning” bet compresses core academics into a short, software-driven block and hands the rest of the day to deeper human work. Strip away the branding and the higher-ed analogue is clear: let AI deliver and drill the content to mastery, and spend scarce, expensive faculty time on the irreplaceable — the seminar, the critique, the mentorship, the live defense. The thing schools keep getting backwards is aiming the technology at the human part and leaving the content delivery to the humans. The reinvention flips it.
And the last piece is the one the brochures will resist: the reinvented college is smaller, more human, and harder to scale — the exact opposite of the MOOC dream of a decade ago. As the synthetic floods the information environment, the premium moves, as the trust section argued, to what cannot be faked at scale — and on a campus that takes the form of a thing built and defended in front of other people. That is good news for what college could be and bad news for its economics, because the high-contact, low-ratio version is the expensive version, and it arrives at the precise moment the demographic cliff and a decade of closures are starving the institutions that would have to deliver it. Which puts the equity problem squarely on the table: if the rebuilt college is a small, in-person, mentor-rich product, it risks becoming one more advantage hoarded at the top — the most selective schools tightening below 4% while everyone else is handed a chatbot and called served. The redesigned model already exists in miniature, and it proves the point both ways: Minerva University runs no lectures at all, only small active-learning seminars — the high-contact form that should rise in value — yet it admits under 2% of applicants, scarcer than any Ivy. The seminar-rich college, so far, is a luxury good. Olah’s “duty to the global poor” lands here too: a reinvention only the wealthy can afford is not a reinvention of college so much as a widening of the gap that college, at its best, once narrowed.
The live test of all this is now running at one of the country’s most tradition-bound universities. In 2026 the University of Chicago gave every student, faculty member, and staffer free enterprise access to Anthropic’s full model line — not just the chatbot but the agentic tools, Claude Code and Cowork, that can read, write, and modify files in a user’s own folders — alongside a $50 million AI research push and faculty committees on AI and education. It is about the most a top university has done on the adoption axis, and putting agentic frontier AI in every student’s hands is genuinely significant — the build-with-it move from the fifth observation, at institutional scale. But notice what it is not: there is no university-wide policy on AI in coursework, each instructor left to permit or prohibit it course by course, and no redesign of how the place assesses or what its degrees certify. Provisioning the tool is the easy half; whether Chicago reinvents anything depends entirely on what it does next with the assignment and the transcript — which is exactly the line between adoption and redesign this section keeps drawing.
The most clear-eyed version of this comes from inside a public research university. Amarda Shehu — the George Mason computer scientist from the fifth observation, now writing as its inaugural chief AI officer, at a public university serving forty thousand students, most of them first-generation — argues that the university has long been telling two incompatible stories: a story of holistic formation to its trustees and the public, and a story of the credential and the return on investment to its students and their families. Prestige and trust held the two in suspension; AI is “the mirror that has shown that both stories cannot be true at the same time.” She is unsparing about the false comfort of a safe-major list — “if a job is a task that can be fully digitized, it is done,” and “anyone who gives you a list is selling you something.” The discipline on the diploma is not the unit of analysis; the formation inside it is. And that formation, she insists, “is not a curriculum. It is an institution” — faculty present to students, advising that persists over years, mentoring that never shows up in a workload calculation — exactly the things universities have been quietly disinvesting in because the loss “did not show up in the metrics.” Her warning carries a hard equity edge: the wealthy privates can extend the hedge another decade, but the reckoning lands on the public access institutions now, and the slow surrender “will be paid for by the students who can least afford it.”
So the crux, stated plainly: reinventing college is not a technology problem and barely a curriculum problem. It is whether institutions built to deliver and certify content can be rebuilt — against their own incentives, their accreditation regimes, and their cost structures — into places that develop and vouch for judgment, that manufacture the apprenticeship the economy stopped supplying, and that do it at a price most students can actually reach. The ones that manage it will be doing something genuinely new. The ones that bolt AI onto the same lecture-and-transcript machine, or retreat into a treadmill of credentials, will go the way of Syracuse — missing targets, staring down deficits, optimizing a product the market has quietly stopped valuing. The deciding variable, here as everywhere in these notes, is not the tool. It is whether the institution rebuilt the work.
17. The backlash is real, and growing
The unease underneath all of this — the students who feel replaceable, the public watching the buildout — does not stay private. It is already taking organized form. When the Future of Life Institute gathered people from the hard right to the far left, labor, the academy, and several faith traditions to find what they could agree on, the result was a cross-ideological “prohuman” declaration whose thirty-three principles — among them that superintelligence should not be built until it can be shown to be safe and controllable, “if ever” — each cleared ninety-percent support; that a group this fractious converged on positions that strong is itself a measure of how fast the worry is consolidating. And it is not only activists and the public who are alarmed: Beth Barnes, who runs METR — the independent outfit the labs invite in to test whether their own models can slip human control — writes plainly that “we are not on top of it”: the field is, she says, likely on track to systems capable of human extinction or permanent disempowerment “quite possibly within the next few years,” its safety teams are “woefully under-resourced,” and any “reasonable civilization” would be moving far more slowly. When the field’s own scorekeeper says that out loud, the backlash starts to look less like technophobia than a rational reading of the gap between pace and preparation. Running alongside all of this is a rising opposition to AI, and schools should not assume they’re insulated from it. The ambivalence is right there in the polling: about half of U.S. adults now use AI chatbots, up from a third in 2024 — yet 63% say the technology is moving too fast and just 16% expect it to have a positive impact on society, with adults under 30 the most skeptical of all. Use is climbing and trust is falling at the same time.
Part of it is about resources. Schools are on tight budgets, and when scarce money goes to an AI chatbot that was thrown at faculty with no real integration, people are right to ask whether it should have been spent elsewhere. Part of it comes from students, who increasingly see the technology as something built to replace them. The environmental argument — strong, in my view — runs underneath all of it, though I’d now narrow where it bites: Amodei dismisses the datacenter water-usage complaint as a non-problem, and on water alone he has a case; the firmer ground is the energy draw and the fights over siting. And increasingly that argument has a physical address: the data center. Even people with little technical understanding of AI experience it as the data center that lands in their community, makes noise, consumes power and water, and, oddly, often arrives with tax breaks that further enrich the industry. Roughly seven in ten Americans (71%, with 48% strongly opposed) say they’d oppose a data center in their community; Georgia is projected to forgo billions in revenue to those incentives; and more than 300 data-center bills were filed across 30-plus states in the first six weeks of 2026 alone. The backlash is bipartisan. And the burden falls unevenly: the centers tend to follow the path the timber and then the coal did before them, into the communities least able to refuse — what Theresa Burriss, from her seat on Virginia’s environmental-justice council, frames as the long habit of treating places like Appalachia as national “sacrifice zones.” Whether a given data center helps or harms a place, she argues, comes down almost entirely to local leadership and whether residents show up to demand transparency — which makes it, at bottom, a civic question.
The opposition also has a violent tail. More than a thousand pages of DHS and FBI reporting, obtained by WIRED and catalogued by Peter Diamandis, describe a rising anti-tech extremism aimed at AI, data centers, and executives — arson at an AI-campus construction site near Paris, a German group claiming sabotage of Berlin power infrastructure over data-center energy demand, and, in April 2026, a twenty-year-old who threw a Molotov cocktail at Sam Altman’s San Francisco home and was charged with attempted murder, carrying a manifesto and a list of AI executives. Firebombing a house is terrorism, not protest — but it is a thermometer. The pressure runs the other way on campus, too: Diamandis describes two pro-AI college students who feel ostracized by classmates for their optimism, called “naive, reckless, greedy, even dangerous.” A reader of his named the honest diagnosis better than he did — the backlash is “a very predictable human response” to lab leaders announcing that millions of livelihoods will vanish in three to five years with no transition plan behind the warning, which makes the anger less technophobia than an accountability the field has not answered. That is the charged ground schools are sending students into, mostly without a word about it.
Wrapped around it is the inequality narrative. The same month a frontier lab’s models made headlines, Tesla shareholders approved a pay package that could make Elon Musk the world’s first trillionaire. To put that in proportion, Amodei reaches for the Gilded Age: John D. Rockefeller’s fortune at its height came to roughly 2% of US GDP — a ratio Musk’s ~$700 billion already exceeds, and that is before most of AI’s economic impact has even landed. Against a backdrop of a widening rich-poor gap and persistent poverty, a lot of people simply see AI as something being hoisted on them for the benefit of those already invested in it — an entire AI world coming for them whether or not they ever choose to use it. Schools have done almost nothing to help students make sense of that feeling.
Hinton offers the frame that turns this from noise into signal. The labs, he notes, like to sell an analogy in which progress is the accelerator and regulation the brake; the better metaphor, he argues, is that regulation is the steering wheel — you can want the car to go fast and still insist it not drive off a cliff. Read that way, the backlash is less a refusal of the technology than a demand to hold the wheel, and teaching students to tell those two things apart is itself a piece of civic education.
18. Radical agency is the only honest answer right now
Let me end the observations on one genuinely hopeful note, because it is real. The same wave that is sawing off the entry-level rung is, at the same moment, the best environment for building something new that has ever existed. Sam Altman puts the leverage concretely: an affordable spend on tokens now lets one person do the work of a 100-person engineering team, something that was simply out of reach a few years ago. Capital is abundant, the tools are cheap and getting cheaper, and an AI-native startup can now recruit talent that the old career ladder would have absorbed. The graduates who can’t find a job and the founders having the best year of their lives are, increasingly, the same people — the difference is whether they are waiting for a rung or making one.
The barbell again — Fei-Fei Li and David Rogier’s map of this same labor market, the top specialist on one end and the high-agency generalist on the other, the middle hollowed out — is, at bottom, an argument about agency. The word they keep returning to is agency, and their most useful move is to pry it loose from the startup world: “entrepreneurial,” Li argues, is just a synonym for agency, as available to a nurse or a teacher as to a founder. That turns the century-old advice — pick a specialty and go deep, which is now eroding except at the very top — into something closer to cultivating the disposition to act.
That is the closest thing I have to a near-term answer for schools. The leverage that used to require a team and a decade now sits inside a laptop, and a student who learns to wield it holds agency on a scale no previous generation of seventeen-year-olds ever did. The instinct to protect students from these tools is understandable and, I think, exactly wrong. The task is to put the tools in their hands and teach them to point them at something worth building — to give them future-shaping power instead of leaving them to absorb future shock, to borrow Peter Diamandis’s reframing: the future happening for them, not to them.
What I keep coming back to
It is not an accident that the Pope reached back 135 years to Rerum Novarum (Leo XIII, 1891). The last time work and human dignity were this unsettled, it was the industrial revolution — and institutions had generations to adapt. This time the same magnitude of change is arriving in semesters.
Underneath all of these observations is one question schools have not seriously asked. If a machine knows more than the teacher and writes better than the student, if cognition itself is now sold by the token, and if college is no longer the scarce prize at the end of the line, then what, exactly, is school for? I don’t think the answer is more content delivery, which is the one thing now in infinite supply. I suspect it lives in the distinctly human capacities no harness can outsource — judgment, argument, deliberation, the ability to stand up and defend a claim in real time — and in the agency to turn these new tools to human ends. It is worth being precise about that first word: Dasey argues schools should retire “critical thinking” — a junk-drawer of analytical sub-skills the machine now does better than we do — in favor of judgment, the contextual, value-laden work of deciding well under uncertainty, the thing medicine and law have always built not through lectures but through authentic practice with real stakes. That is the same move the redesign in the first observation asks of every assignment, and it may be the truest answer to what the years are for.
Even the optimists circle the same ground: Bill Gates, no pessimist about technology, says the “demand for people who help others… will never go away.” Even Peter Diamandis, whose maximalist “supersonic tsunami” forecast promises ASI by 2031 and a quadrillion-dollar economy, gives students the same instruction these notes do: stop memorizing what a machine knows better, and master “taste, judgment, the ability to ask the right question and rally humans around an answer” — “learn to direct intelligence, not compete with it.” The builders themselves now describe the human residual in nearly the terms I would: Altman and Pachocki, in OpenAI’s own statement of plans, insist that “entirely automating everything is not the future we want,” that as the systems grow more capable “the human role becomes more important — setting direction, making tradeoffs, applying judgment,” and that the lasting human job is “deciding what is worth doing.” Coming from the company racing hardest to automate everything, that is either a genuine concession or a convenient one — yet on the substance it is the same ground these notes keep returning to. But that is the subject of the notes still to come.
For now, the honest summary of the 2025–26 school year is this: the ground moved, most schools felt it, and very few have decided what to do about it. And every year a school spends deciding is a year its students learn less — which means the deciding is not neutral, and the clock is not paused.
The world of 2026-7 will not be anything like the world of 2025-6, let alone 2022-3.
What you are describing here in the domain of assessment philosophy is the difference between assessing intelligence and assessing intellectual virtue. I plan to post an article on my Sub that references your expository here. Well done, Stefan.
Thank you for saying out loud so many of the things I've been privately seething about my entire career in higher ed!
side note: The image at the top of the article says 2015-16 instead of 2025-26.