Debate Cards: Teaching Students to Think With Evidence

And samples from the key issues in AI

Join us for the GlobalAI Debates!

If you’ve never encountered competitive debate, you may not know about one of its most powerful learning tools: the debate card.

A debate card is a structured excerpt from a source, formatted so that a debater can quickly deploy evidence in a live argument. Every card has four parts:

Tag — A one-sentence summary of the argument the evidence supports. This is written by the student in their own words. For example: “AI-driven automation will displace 300 million jobs globally by 2030.”

Citation — The author, their qualifications, the source, and the date. Knowing who said something and when matters. A 2024 McKinsey report carries different weight than an anonymous blog post.

Card body — The full text of the relevant passage from the source.

Underlining — The key sentences within the card body that the debater actually reads aloud in a round. The rest is there for context, but the underlined portions are the core claim. In written form, these are usually bolded or underlined.

The format looks something like this:

Autonomous weapons development is accelerating without meaningful oversight.

Scharre, 25 — Paul Scharre, VP and Director of Studies at the Center for a New American Security, former DoD policy advisor, Foreign Affairs, February 2025.

The rapid pace of military AI development has outstripped the ability of existing legal frameworks to regulate autonomous weapons systems. Multiple nations are now fielding drones and autonomous targeting systems with decreasing levels of human oversight, and no binding international agreement governs their use. The gap between capability and governance continues to widen as defense departments prioritize speed over deliberation.

The debater would read the bolded portions. The rest provides context if challenged.

Why Cards Matter

Cards aren’t just a debate convention. They’re a learning method disguised as competition prep.

They force students to read. Not skim, not summarize from an AI-generated overview — actually find, select, and engage with primary and secondary sources. A student building a file of cards on AI policy has read dozens of articles, reports, and book chapters closely enough to identify the specific passages that matter.

They force close reading. Choosing what to underline requires judgment. Which two sentences carry the argument? What’s context and what’s the claim? This is analytical reading at a high level.

They force engagement with opposing evidence. In debate, your opponent has cards too. You learn that smart people with good evidence disagree with you — and you have to respond to their best arguments, not their worst. This is the opposite of how most people encounter disagreement online.

They build cumulative knowledge. Cards support each other. A file on a topic might contain thirty cards that construct an argument across economics, ethics, technology, and governance. Students learn to think in systems, not soundbites.

Do you Want Your Students to Read and Think?

Educators have been complaining that students are just using AI to complete their assignments and they aren’t thinking anymore. Well, try giving your students a debate card assignment.

Building Cards from Sources. Give students 2–3 articles on one of the topics — job displacement, autonomous weapons, surveillance, AI personhood — and have them produce 5 cards. They choose the tags, select the passages, decide what to underline. This forces them to read the full articles and make judgment calls about what matters.

Card vs. Card. Pair students up. One has cards arguing AI job displacement will be catastrophic; the other has cards arguing new jobs will emerge. They have to respond to each other’s specific evidence, not just assert opinions. This teaches that disagreement lives in the evidence, not in volume.

Find the Weak Card. Give students a set of 10 cards on a topic where 2–3 have weak sources, outdated citations, or tags that overstate what the evidence actually says. Students have to identify which cards they’d cut from their file and explain why. This builds source evaluation skills without a lecture on “media literacy.”

Tag Rewriting. Provide cards with the body and citation but no tag. Students write their own. Then compare — did different students read the same passage and frame the argument differently? This surfaces how framing shapes persuasion even when the evidence is identical.

Build the File. Assign a topic — say, mass surveillance and AI — and have students build a 15-card file over two weeks. They need cards on both sides. At the end, they write a one-page assessment of which side has stronger evidence and why. The constraint of finding evidence for both sides prevents the assignment from becoming a book report that confirms what they already believe.

The Missing Question. Give students your list of ignored topics (job loss, surveillance, autonomous weapons, AI rights, schools ignoring the new world). Have them pick one, build 5 cards, and then write a one-paragraph argument for why their school’s curriculum should address it. This turns the meta-argument from your article into something students own.

Update the Card. Give students a card from 2023 and ask them to find a more recent source that either strengthens, weakens, or overtakes the original evidence. This teaches that evidence has a shelf life — especially in AI, where six months can be a lifetime.

The Oral Drill. Students pick their three strongest cards on a topic and deliver them aloud in 90 seconds — reading only the tag and underlined portions. Listeners have to identify the core claim from what was read. This builds both public speaking and listening comprehension under pressure.

How Cards Actually Promote Thinking

The most common objection to debate cards is that they look like rote work — find a quote, copy it, move on. But the thinking happens at every step, and it’s the kind of thinking that’s hardest to teach any other way.

Choosing what to cut is an argument. When a student reads a 3,000-word article and selects 150 words for their card, they’ve made a claim about what matters. They’ve decided what the author’s strongest point is, which evidence is most persuasive, and what can be left out. That’s not copying — that’s analysis.

Writing the tag is original reasoning. The tag isn’t a summary of the article. It’s the student’s assertion about what the evidence proves. Two students can read the same source and write completely different tags depending on what argument they’re building. The tag is where the student’s own thinking lives.

Underlining forces prioritization under constraint. You can’t read everything in a debate round — you have maybe 15 seconds per card. So the student has to decide: which two or three sentences carry the full weight of this argument? That’s a judgment call that requires understanding not just what the author said but what will be most compelling to a listener hearing it once, in real time.

Cards don’t exist alone — they exist in relationship to other cards. Building a file means constructing a coherent argument across multiple sources. This card establishes the problem, that one quantifies it, the next one answers the most likely objection. Students are doing what professional researchers do: synthesizing across sources to build something larger than any single piece of evidence.

The format makes thinking visible and contestable. When your argument is a card, your opponent can challenge the source, question whether the tag accurately represents the evidence, or read a different portion of the same article that contradicts your underline. You can’t hide behind vague assertions. Everything is on the table — and that’s what forces intellectual honesty.

The irony is that in an era when AI can generate a passable essay in seconds, the debate card is more valuable than ever. You can’t fake the judgment involved in selecting, tagging, and underlining evidence. The thinking is the format.

The deepest thinking happens when someone else’s card is on the table and you have to decide, in real time, what’s wrong with it.

When a debater hears an opponent read a card, they can’t just disagree — they have to figure out why the evidence fails. That forces a rapid series of intellectual moves that no essay assignment replicates.

Is the source credible? A card from a think tank funded by defense contractors arguing autonomous weapons are safe carries different weight than one from an independent researcher. Students learn to evaluate not just what was said but who said it and what incentives shaped the claim. This isn’t a worksheet on “identifying bias” — it’s a live moment where getting it wrong means losing the round.

Does the tag match the evidence? One of the most common and devastating moves in debate is pointing out that an opponent’s tag overstates what the card actually says. The tag claims “AI will eliminate 300 million jobs” but the underlined evidence says “up to 300 million jobs could be exposed to automation.” That gap between claim and evidence is where intellectual dishonesty lives, and debaters learn to spot it — in their opponents’ cards and, eventually, in their own.

Is the evidence outdated or overtaken? A card from 2023 about AI capabilities may be functionally obsolete. A student who can say “that was written before autonomous research agents completed 126 experiments overnight — the landscape has changed” is doing something more sophisticated than rebuttal. They’re reasoning about the shelf life of knowledge itself.

What does the card leave out? Every card is a selection — 150 words pulled from a 3,000-word article. What got cut? Maybe the next paragraph contained a qualification that undermines the whole argument. Debaters learn to ask what’s missing, which is the single most important critical thinking skill there is. Most persuasion works by omission, and debate trains you to notice it.

Can you turn the evidence? Sometimes the best response to a card isn’t to deny it but to accept it and show it actually supports your side. Your opponent reads evidence that AI is automating jobs at unprecedented speed — and you argue that’s exactly why we need the policy you’re proposing. This requires holding two interpretations of the same evidence in your head simultaneously and arguing for the one your opponent didn’t intend. That’s high-level reasoning.

What makes all of this different from a classroom discussion is that there are consequences. You’re not sharing your opinion in a circle where every answer gets a nod. You’re making a claim with evidence, and someone is actively trying to dismantle it while you listen. That pressure — respectful, structured, evidence-based pressure — is what turns reading into thinking and thinking into judgment.

Cards on the Questions Nobody Is Asking

I’ve been putting together evidence files on issues that the current AI conversation is largely ignoring. While schools debate whether students should be allowed to use ChatGPT, the world is moving on to questions with far higher stakes.

Here are a few of the topics I have cards on —

We need AI debates to guide society through the AI transition/revolution

Dwarkesh Patel, March 11, 2026 American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

The only way we can preserve our free society is if we make laws and norms through our political system that it is unacceptable for the government to use AI to enforce mass surveillance and censorship and control. Just as after WW2, the world set the norm that it is unacceptable to use nuclear weapons to wage war. I want to be clear: these are extremely confusing and difficult questions to think about. I kept changing my mind back and forth on many of them in the process of writing this essay. I reserve the right to change my mind again in the future. In fact, I think it’s essential to change our minds as AI progresses and we learn more. That’s the whole point of conversation and debate. Someday people will look back on this period the way we look back on the Enlightenment. People having big important debates right as the world was about to undergo these massive technological, social, and political revolutions. And some of these thinkers actually managed to get a couple of the big things right, for which we are now the beneficiaries. We owe it to our future to at least attempt to think through these new questions raised by AI.

Anthropic, March 11, 2026, https://www.anthropic.com/news/the-anthropic-institute, Introducing The Anthropic Institute

One of our company’s core convictions is that AI development is accelerating: that the improvements we make are compounding over time. Because of this, extremely powerful AI, like the kind our CEO Dario Amodei describes in Machines of Loving Grace, is coming far sooner than many think.

If this is right, society is shortly going to need to confront many massive challenges. How will powerful AI systems reshape our jobs and economies? What kinds of opportunities for greater societal resilience will they give us? What kinds of threats will they magnify or introduce? What are the expressed “values” of AI systems and how will society help companies determine what the appropriate values are? And, if the recursive self-improvement of AI systems does begin to occur, who in the world should be made aware, and how should these systems be governed?

Students not prepared to succeed in the AI World

Ted Dintersmith, founder of What School Could Be, March 13, 2026, https://www.linkedin.com/feed/update/urn:li:activity:7438199724261888000/?originTrackingId=pHGyBFr4p7%2BTiG1OwjF5Dg%3D%3D

Each year, I guest-teach a session at an elite university. Last month, I asked about 100 juniors and seniors, “How many of you are proficient today at using AI to make yourself and those around you far more productive?” Not one hand went up. A bit stunned, I pushed further. Still no hands, perhaps out of fear of being flagged as a potential cheater. I explained that AI proficiency is a total game-changer -- opening career doors and fueling their community service activities. Their very consequential choice -- focus on getting that GPA up a smidgeon, or master the use of AI.

Government is not qualified or more ready to regulate superintelligence than private companies

Dwarkesh Patel, March 11, 2026 American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

While some regulation might be inevitable, I think it’d be a terrible idea or the government to wholesale take over this technology. Ben Thompson had a post last Monday where he made the point that people like Dario have compared the technology they’re developing to nuclear weapons - specifically in the context of the catastrophic risk it poses, and why we need to export control it from China. But then you oughta think about what that logic implies: “if nuclear weapons were developed by a private company, and that private company sought to dictate terms to the U.S. military, the U.S. would absolutely be incentivized to destroy that company.” And honestly, safety aligned people have actually made similar arguments. Leopold Aschenbrenner, who is a former guest and a good friend, wrote in his 2024 Situational Awareness memo, “I find it an insane proposition that the US government will let a random SF startup develop superintelligence. Imagine if we had developed atomic bombs by letting Uber just improvise.”

And my response to Leopold’s argument at the time, and Ben’s argument now, is that while they’re right that it’s crazy that we’re entrusting private companies with the development of this world historical technology, I just don’t see the reason to think that it’s an improvement to give this authority to the government. Nobody is qualified to steward the development of superintelligence. It is a terrifying, unprecedented thing that our species is doing right now, and the fact that private companies aren’t the ideal institutions to take up this task does not mean the Pentagon or the White House is.

AIs will not resist immoral actions the way people do, alignment with human direction risks mass death

Dwarkesh Patel, March 11, 2026 American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic Alignment - to whom?

And this gets us to the issue of alignment. What I have just described to you - an army of extremely obedient employees - is what it would look like if alignment succeeded - that is, we figured out at a technical level how to get AI systems to follow someone’s intentions. And the reason it sounds scary when I put it in terms of mass surveillance or robot armies is that there is a very important question at the heart of alignment which we just haven’t discussed much as a society. Because up till now, AIs were just capable enough to make the question relevant: to whom or what should the AIs be aligned? In what situations should the AI defer to the end user versus the model company versus the law versus its own sense of morality?

This is maybe the most important question about what happens with powerful AI systems. And we barely talk about it. It’s understandable why we don’t hear much about it. If you’re a model company, you don’t really wanna be advertising that you have complete control over a document that determines the preferences and character of what will eventually be almost the entire labor force, not just for private sector companies, but also for the military and the civilian government.

We’re getting to see, with this DoW/Anthropic spat, a much earlier version of the highest stakes negotiations in history. By the way, make no mistake about it - with real AGI the stakes are even much higher than mass surveillance. This is just the example that has come up already relatively early on in the development of AGI.

The military insists that the law already prohibits mass surveillance, and so Anthropic should agree to let their models be used for “all lawful purposes”. Of course, as we saw from the 2013 Snowden revelations, even in this specific example of mass surveillance , the government has shown that it will use secret and deceptive interpretations of the law to justify its actions. Remember, what we learned from Snowden was that the NSA, which, by the way, is part of the Department of War, used the 2001 Patriot Act’s authorization to collect any records “relevant” to an investigation to justify collecting literally every phone record in America. The argument went that it was all “relevant” because some subset might prove useful in some future investigation. They ran this program for years under secret court approval. So when the Pentagon today says, “We would never use AI for mass surveillance, it’s already illegal, your red lines are unnecessary”, it would be extremely naive to take that at face value. No government is going to call its own actions “mass surveillance”. For the government, it will always have a different label.

So then Anthropic comes back and says, “No, we want red lines separate from ‘all lawful purposes,’ and we want the right to refuse you service when we believe those red lines are being violated.” But think about it from the military’s perspective. In the future, almost every soldier in the field, and every bureaucrat and analyst and even general in the Pentagon, is going to be an AI. And that AI is, on current track, going to be supplied by a private company. I’m guessing Hegseth is not thinking about “genAI” in those terms just yet. But sooner or later, it will be obvious to everyone what the stakes here are, just as after 1945, the strategic importance of nuclear weapons became clear to everyone. And now the private company insists that it reserves the right to say, “Hey, Pentagon, you’re breaking the values we embedded in our contract, so we’re cutting you off.”

Maybe in the future, Claude will have its own sense of right and wrong, and it will be smart enough to just personally decide that it’s being used against its values. For the military, maybe that’s even scarier. I’ll admit that at first glance, “let the AI follow its own values” sounds like the pitch for every sci-fi dystopia ever made. The Terminator has its own values. Isn’t this literally what misalignment is? But I think situations like this actually illustrate why it matters that AIs have their own robust sense of morality. Some of the biggest catastrophes in history were avoided because the boots on the ground refused to follow orders. One night in 1989, the Berlin Wall fell, and as a result, the totalitarian East German regime collapsed, because the guards at the border refused to shoot down their fellow countrymen who were trying to escape to freedom. Maybe the best example is Stanislav Petrov, who was a Soviet lieutenant colonel on duty at a nuclear early warning station. His sensors reported that the United States had launched five intercontinental continental ballistic missiles into the Soviet Union. But he judged it to be a false alarm, and so he broke protocol and refused to alert his higher-ups. If he hadn’t, the Soviet higher-ups would likely have retaliated, and hundreds of millions of people would have died.

Of course, the problem is that one person’s virtue is another person’s misalignment. Who gets to decide what moral convictions these AIs should have - in whose service they may even decide to break the chain of command? Who gets to write this model constitution that will shape the characters of the intelligent, powerful entities that will operate our civilization in the future?

I like the idea that Dario laid out when he came on my podcast: different AI companies can build their models using different constitutions, and we as end users can pick the one that best achieves and represents what we want out of these systems. I think it’s very dangerous for the government to be mandating what values AIs should have.

The US Department of War will undermine AI that has different values and resists

Ashley Kapoot, March 12, 2026, CNBC, Anthropic’s Claude would ‘pollute’ defense supply chain: Pentagon CTO, https://www.cnbc.com/2026/03/12/anthropic-claude-emil-michael-defense.html

Defense Department CTO Emil Michael said Anthropic’s Claude AI models would “pollute” the agency’s supply chain because they have “a different policy preference” that is baked in. Anthropic has sued the Trump administration to reverse the supply chain risk designation and said the government is putting hundreds of millions in contracts in doubt. “This is not meant to be punitive,” Michael told CNBC’s “Squawk Box.” Defense Undersecretary Emil Michael: Anthropic’s Claude would ‘pollute’ defense supply chainwatch now VIDEO08:50 Defense Undersecretary Emil Michael: Anthropic’s Claude would ‘pollute’ defense supply chain Defense Department CTO Emil Michael on Thursday said Anthropic’s Claude artificial intelligence models would “pollute” the agency’s supply chain because they have “a different policy preference” that is baked in. “We can’t have a company that has a different policy preference that is baked into the model through its constitution, its soul, its policy preferences, pollute the supply chain so our warfighters are getting ineffective weapons, ineffective body armor, ineffective protection,” Michael told CNBC’s “Squawk Box.” “That’s really where the supply chain risk designation came from.” Anthropic is the first American company to publicly be labeled a supply chain risk, an extraordinary move that’s historically been reserved for foreign adversaries. The designation will require defense contractors and vendors to certify that they don’t use Claude in their work with the Pentagon. Michael’s comments on Thursday are the clearest explanation the DOD has offered about why it believes Anthropic is a supply chain risk. The agency sent an official letter to notify the company about the designation earlier this month, but the letter did not outline what risk Claude poses to national security. Anthropic sued the Trump administration on Monday, calling the government’s actions “unprecedented and unlawful.” Anthropic said in a filing that the company was being harmed “irreparably,” and that hundreds of millions of dollars worth of contracts are in jeopardy. “This is not meant to be punitive,” Michael said Thursday. He added that Anthropic has a “huge commercial business,” and that a “tiny fraction” comes from the U.S. government. Michael also dismissed Anthropic’s claim that the government has actively reached out to companies and told them them not to use Anthropic, calling the notion “rumors.” “The Department of War is not reaching out to companies to tell them what to do, so long as it’s not in our supply chain,” he said. Read more CNBC tech news Sam Altman faced ‘serious questions’ in meeting with lawmakers about OpenAI’s defense work Adobe CEO Shantanu Narayen says he will step down after company installs successor Anthropic’s Claude would ‘pollute’ defense supply chain: Pentagon CTO Palantir is still using Anthropic’s Claude as Pentagon blacklist plays out, CEO Karp says Anthropic was founded in 2021 by a group of researchers and executives who defected from OpenAI. The company is best known for its family of Claude models, and it’s had early success selling into large enterprises, including the DOD. The startup has drafted and published a “constitution” that it uses to help train its mainline, general-access Claude models. Anthropic said the constitution plays a “crucial role” in this process, and that its content “directly shapes Claude’s behavior,” according to its website. Anthropic shared the most recent version of Claude’s constitution in January. “In it, we explain what we think it means for Claude to be helpful while remaining broadly safe, ethical, and compliant with our guidelines,” Anthropic said in a blog post. “The constitution gives Claude information about its situation and offers advice for how to deal with difficult situations and tradeoffs, like balancing honesty with compassion and the protection of sensitive information.”

99% of labor will be AI

Dwarkesh Patel, March 11, 2026 American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

By now, I’m sure you’ve heard that the Department of War has declared Anthropic a supply chain risk, because Anthropic refused to remove redlines around the use of their models for mass surveillance and for autonomous weapons.

Honestly I think this situation is a warning shot. Right now, LLMs are probably not being used in mission critical ways. But within 20 years, 99% of the workforce in the military, the government, and the private sector will be AIs. This includes the soldiers (by which I mean the robot armies), the superhumanly intelligent advisors and engineers, the police, you name it.

We are on the threshold of AGI; it’s on the horizon

Note: They use debate to get to the best answer 

Dr. Demis Hassabis, Google Deep Mind, March 10, 2026, https://deepmind.google/blog/10-years-of-alphago/, From games to biology and beyond: 10 years of AlphaGo’s impact

Gemini, our largest and most capable model, recently went even further. An advanced version of its Deep Think mode achieved gold-medal level performance at the 2025 IMO using an approach inspired by AlphaGo. Since then, Deep Think has been applied to even more complex, open-ended challenges across science and engineering. Algorithm discovery: Just as AlphaGo searched for the best move in a game, our coding agent AlphaEvolve explores the space of computer code to discover more efficient algorithms. It had its own Move 37 moment when it found a novel way to multiply matrices, a fundamental mathematical operation powering nearly all modern neural networks. AlphaEvolve is now being tested on problems ranging from data center optimization to quantum computing. Scientific collaboration: We are integrating the search and reasoning principles pioneered with AlphaGo into an AI co-scientist. By having agents ‘debate’ scientific ideas and hypotheses, this system acts as a collaborator capable of performing the rigorous thinking necessary to identify patterns in data and solve sophisticated problems. In validation studies at Imperial College London, it analyzed decades of literature and independently arrived at the same hypothesis about antimicrobial resistance that researchers had spent years developing and validating experimentally. We’ve also used AI to better understand the genome, advance fusion energy research, improve weather prediction and more. As impressive as our scientific models are, they are highly specialized. To achieve fundamental breakthroughs like creating limitless clean energy or solving diseases that we don’t understand today, we need general AI systems that can find underlying structure and connections between different subject areas, and help us to come up with new hypotheses like the best scientists do. Future of intelligence For an AI to be truly general, it needs to understand the physical world. We built Gemini to be multimodal from the beginning so it could understand not just language, but also audio, video, images and code to build a model of the world. To think and reason across these modalities, the latest Gemini models use some of the techniques we pioneered with AlphaGo and AlphaZero. The next generation of AI systems will also need to be able to call upon specialized tools. For example, if a model needed to know the structure of a protein it could use AlphaFold for that. We think the combination of Gemini’s world models, AlphaGo’s search and planning techniques, and specialized AI tool use will prove to be critical for AGI. True creativity is a key capability that such an AGI system would need to exhibit. Move 37 was a glimpse of AI’s potential to think outside the box, but true original invention will require something more. It would need to not only come up with a novel Go strategy, as AlphaGo impressively did, but actually invent a game as deep and elegant, and as worthy of study as Go. Ten years after AlphaGo’s legendary victory, our ultimate goal is on the horizon. The creative spark first seen in Move 37 catalyzed breakthroughs that are now converging to pave the path towards AGI - and usher in a new golden age of scientific discovery.

Mass surveillance is legal; the government can use LLMs to organize all the legal data in a way that has never been done

D Dwarkesh Patel, March 11, 2026, American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

Mass surveillance is, at least in certain forms, legal. It just has been impractical so far. Under current law, you have no Fourth Amendment protection over data you share with a third party, including your bank, your phone carrier, your ISP, and your email provider. The government reserves the right to purchase and obtain and read this data in bulk without a warrant.

What’s been missing is the ability to actually do anything with all of this data — no agency has the manpower to monitor every camera feed, cross-reference every transaction, or read every message. But that bottleneck goes away with AI.

There are 100 million CCTV cameras in America. You can get pretty good open source multimodal models for 10 cents per million input tokens. So if you process a frame every ten seconds, and each frame is 1,000 tokens, you’re looking at a yearly cost of about 30 billion dollars to process every single camera in America. And remember that a given level of AI ability gets 10x cheaper year over year - so a year from now it’ll cost 3 billion, and then a year after 300 million, and by 2030, it might be cheaper for the government to be able to understand what is going on in every single nook and cranny of this country than it is to remodel the White House.

Once the technical capacity for mass surveillance and political suppression exists, the only thing standing between us and an authoritarian surveillance state is the political expectation that this is not something we do here. And this is why I think what Anthropic did here is so valuable and commendable, because it is helping set that norm and precedent.

AI generated deep fakes result in death

Mahsa Alimardani, March 13, 2023, The Atlantic, The Fog of AI, The Fog of AI, https://www.theatlantic.com/ideas/2026/03/ai-imagery-iran-war/686347/

In February 27, an AI-generated image appeared on Instagram purporting to show heavy military equipment stationed inside Karimian Elementary School in Isfahan, Iran. The post, shared by accounts including the Free Union of Iranian Workers, an independent labor union operating inside Iran whose leaders have been jailed by the regime, read: “This is not a military zone! It’s Karimian Elementary.” The image carried a visible Google Gemini watermark, indicating that it had been created by the software. The school posted a rebuttal, noting that the equipment could not physically fit on the premises. Iranian-diaspora fact-checkers confirmed that the image was fabricated.

The next day, Shajareh Tayyebeh, a girls’ elementary school in the southern city of Minab, was hit in the first wave of strikes on Iran. Iranian authorities reported at least 175 people dead, many of them children. The exact death toll has not been independently confirmed, but a New York Times investigation verified that the school had been hit by a precision strike at the same time as attacks on an adjacent naval base, and a preliminary investigation by the American military concluded that U.S. forces were most likely responsible. The school sat on the grounds of the Iranian navy’s Asef Brigade barracks, an active military base. The building had been converted from military use, and served children from military and civilian families.

In short: The day before the strikes began, an AI image on social media planted the notion that the regime hides military equipment in schools. The next day, a real school—once part of a military compound but walled off from it since 2016, according to Human Rights Watch—was destroyed. The fake was wrong about Karimian, but by the time the Minab strike happened, audiences were primed to believe that a school was a legitimate military target, not the site of a civilian catastrophe. Layer by layer, an accumulation of AI imagery circulated on social media that made it difficult to establish what happened to these children. This is the fog that AI has introduced to the war in Iran. This isn’t a war where AI fakes fool everyone nor where detection tools catch everything. We live in a world where real photographs of real civilian deaths are called fake, and where fake images are used to illustrate real deaths. Where correct identification of one fake image is used to cast doubt on real images, where incorrect detection is authoritative, and where all of it happens faster than any institution, newsroom, fact-checker, photo wire service, or platform can process. The fog of AI does not need every piece of content to be fabricated. It needs the question Is this real? to become close to unanswerable.

AI structurally favors mass surveillance and empowers mass violence

Dwarkesh Patel, March 11, American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

What we’re learning from this episode is that the government actually has way more leverage over private companies than we realized. Even if this supply chain restriction is backtracked (which prediction markets currently give it a 81% chance of happening), the President has so many different ways in which he can make your life difficult if you’re a company that is resisting him. The federal government controls permitting for new power generation, which is needed for datacenters. It oversees antitrust enforcement. The federal government has contracts with all the other big tech companies whom Anthropic needs to partner with for chips and for funding - and they could make it an unspoken condition for such contracts that those companies can no longer do business with Anthropic. People have proposed that the real problem here is that there’s only 3 leading AI companies. This creates a clear and narrow target for the government to apply leverage on in order to get what they want out of this technology. But if there’s wide diffusion, then from the government’s perspective, the situation is even easier. Maybe the best models of early 2027 (if you engineered the safeguards out) - the Claude 6 and Gemini 5 - will be capable of enabling mass surveillance. But by late 2027, and certainly by 2028, there will be open source models that do the same thing. So in 2028, the government can just say, “Oh Anthropic, Google, OpenAI, you’re drawing a line in the sand? No issue - I’ll just run some open source model that might not be at the frontier, but is definitely smart enough to note-take a camera feed.” The more fundamental problem is just that even if the three leading companies draw lines in the sand, and are even willing to get destroyed in order to preserve those lines, it doesn’t really change the fact that the technology itself is just a big boon to mass surveillance and control over the population. And so then the question is, what do we do about it?

Honestly, I don’t have an answer. You’d hope there’s some symmetric property of the technology — some way we as citizens can use AI to check government power as effectively as the government can use AI to monitor and control its population. But realistically, I just don’t think that’s how it’s going to shake out. You can think of AI as giving everybody more leverage on whatever assets and authority they currently have. And the government is already starting with a monopoly of violence. Which they can now supercharge with extremely obedient employees that will not question the government’s orders.

Vague AI regulations enable abuse of government power and manipulation of the citizenry

Dwarkesh Patel, March 11, 2026 American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

I cannot imagine how a regulatory framework built around the concepts that underlie AI risk discourse will not be abused by wannabe despots - the underlying terms are so vague and open to interpretation that you’re just handing a power hungry leader a fully loaded bazooka. ‘Catastrophic risk.’ ‘Mass persuasion risk.’ ‘Threats to national security.’ ‘Autonomy risk.’ These can mean whatever the government wants them to mean. Have you built a model that tells users the administration’s tariff policy is misguided? That’s a deceptive, manipulative model — can’t deploy it. Have you built a model that refuses to assist with mass surveillance? That’s a threat to national security. In fact, the government may say, you’re not allowed to build any model which is trained to have its own sense of right and wrong, where it refuses government requests which it thinks cross a redline - for example, enabling mass surveillance, prosecuting political enemies, disobeying military orders that break the US constitution - because that’s an autonomy risk!

Look at what the current government is already doing in abusing statutes that have nothing to do with AI to coerce AI companies to drop their redlines on mass surveillance. The Pentagon had threatened Anthropic with two separate legal instruments. One was a supply chain risk designation — an authority from the 2018 defense bill meant to keep Huawei components out of American military hardware. The other was the Defense Production Act — a statute passed in 1950 so that Harry Truman could keep steel mills and ammunition factories running during the Korean War.

Do you really want to hand the same government a purpose-built regulatory apparatus on AI - which is to say, directly at the thing the government will most want to control? I know I’ve repeated myself here 10 times, but it is hard to emphasize how much AI will be the substrate of our future civilization. You and I, as private citizens, will have our access to all commercial activity, to information about what is happening in the world, to advice about what we should do as voters and capital holders, mediated through AIs. Mass surveillance, while very scary, is like the 10th scariest thing the government could do with control over the AI systems with which we will interface with the world.

A company will inevitably sell to the government and enable mass surveillance

Dwarkesh Patel, March 11, 2026, American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

If in the future that stops being the case - if only one entity ends up being capable of building the robot armies and the superhuman hackers, and we had reason to worry that they could take over the whole world with their insurmountable lead, then I agree – it would not be acceptable to have that entity be a private company. And so honestly, I think my crux against the people who say that because AI is so powerful we cannot allow it to be shaped by private hands is that I just expect this technology to be much more multi-polar than they do, with lots of competitive companies at each layer of the supply chain.

And it is for this reason that unfortunately, individual acts of corporate courage will not solve the problem we are faced with here, which is just that structurally AI favors authoritarian applications, mass surveillance being one among many. Even if Anthropic refuses to have its models be used for such uses, and even if the next two frontier labs do the same, within 12 months everyone and their mother will be able to train AIs as good as today’s frontier. And at that point, there will be some AI vendor who is capable and willing to help the government enable mass surveillance.

We shouldn’t give the government control over an entire technology type

Dwarkesh Patel, March 11, 2026 American commentator and podcaster, The most important question nobody’s asking about AI, https://www.dwarkesh.com/p/dow-anthropic

Yes - if a single private company were the only entity capable of building nuclear weapons, the government would not tolerate that company claiming veto power over how those weapons were used. I think this nuclear weapons analogy is not the correct way to think about AI. For at least two important reasons:

First, AI is not some self-contained pure weapon. A nuclear bomb does one thing. AI is closer to the process of industrialization itself — a general-purpose transformation of the economy with thousands of applications across every sector. If you applied Thompson’s or Aschenbrenner’s logic to the industrial revolution — which was also, by any measure, world-historically important — it would imply the government had the right to requisition any factory, dictate terms to any manufacturer, and destroy any business that refused to comply. That’s not how free societies handled industrialization, and it shouldn’t be how they handle AI.

People will say, “Well, AI will develop unprecedentedly powerful weapons - superhuman hackers, superhuman bioweapons researchers, fully autonomous robot armies, etc - and we can’t have private companies developing that kind of tech.” But the Industrial Revolution also enabled new weaponry that was far beyond the understanding and capacity of, say, 17th century Europe - we got aerial bombardment, and chemical weapons, not to mention nukes themselves. The way we’ve accommodated these dangerous new consequences of modernity is not by giving the government absolute control over the whole industrial revolution (that is, over modern civilization itself), but rather by coming up with bans and regulations on those specific weaponizable use cases. And we should regulate AI in a similar way - that is, ban specific destructive end uses (which would also be unacceptable if performed by a human - for example, launching cyber attacks). And there should also be laws which regulate how the government might abuse this technology. For example, by building an AI-powered surveillance state.

AGI critical to advances in medicine and science

Dr. Demis Hassabis, Google Deep Mind, March 10, 2026, https://deepmind.google/blog/10-years-of-alphago/, From games to biology and beyond: 10 years of AlphaGo’s impact

I believe the greatest lesson AlphaGo offered was a definitive preview of the AI era—proving it wasn’t some distant, vague future, but a reality arriving on our doorstep. It served as a “roadmap from the future,” sending a clear signal to humanity about how the world was about to change.

By proving it could navigate the massive search space of a Go board, AlphaGo demonstrated the potential for AI to help us better understand the vast complexities of the physical world. We started by attempting to solve the protein folding problem, a 50-year grand challenge of predicting the 3D structure of proteins - information that is crucial for understanding diseases and developing new drugs.

In 2020, we finally cracked this longstanding scientific problem with our AlphaFold 2 system. From there, we folded the structures for all 200 million proteins known to science and made them freely available to scientists in an open-source database. Today, over 3 million researchers around the world use the AlphaFold database to accelerate their important work on everything from malaria vaccines to plastic-eating enzymes. And in 2024, it was the honor of a lifetime for John Jumper and I to be awarded the Nobel Prize in Chemistry for leading this project, on behalf of the entire AlphaFold team.

Since AlphaGo’s win, we’ve applied its groundbreaking approach to many other areas of science and mathematics, including:

Mathematical reasoning: The most direct descendant of AlphaGo’s architecture, AlphaProof learned to prove formal mathematical statements using a combination of language models and AlphaZero’s reinforcement learning and search algorithms. Alongside AlphaGeometry 2, it became the first system to achieve a medal-standard (silver) at the International Mathematical Olympiad (IMO), proving AlphaGo’s methods could unlock advanced mathematical reasoning and laying the foundation for our most capable general models.

Gemini, our largest and most capable model, recently went even further. An advanced version of its Deep Think mode achieved gold-medal level performance at the 2025 IMO using an approach inspired by AlphaGo. Since then, Deep Think has been applied to even more complex, open-ended challenges across science and engineering.

Algorithm discovery: Just as AlphaGo searched for the best move in a game, our coding agent AlphaEvolve explores the space of computer code to discover more efficient algorithms. It had its own Move 37 moment when it found a novel way to multiply matrices, a fundamental mathematical operation powering nearly all modern neural networks. AlphaEvolve is now being tested on problems ranging from data center optimization to quantum computing.

Scientific collaboration: We are integrating the search and reasoning principles pioneered with AlphaGo into an AI co-scientist. By having agents ‘debate’ scientific ideas and hypotheses, this system acts as a collaborator capable of performing the rigorous thinking necessary to identify patterns in data and solve sophisticated problems. In validation studies at Imperial College London, it analyzed decades of literature and independently arrived at the same hypothesis about antimicrobial resistance that researchers had spent years developing and validating experimentally.

We’ve also used AI to better understand the genome, advance fusion energy research, improve weather prediction and more.

As impressive as our scientific models are, they are highly specialized. To achieve fundamental breakthroughs like creating limitless clean energy or solving diseases that we don’t understand today, we need general AI systems that can find underlying structure and connections between different subject areas, and help us to come up with new hypotheses like the best scientists do.

Future of intelligence

For an AI to be truly general, it needs to understand the physical world. We built Gemini to be multimodal from the beginning so it could understand not just language, but also audio, video, images and code to build a model of the world.

To think and reason across these modalities, the latest Gemini models use some of the techniques we pioneered with AlphaGo and AlphaZero.

The next generation of AI systems will also need to be able to call upon specialized tools. For example, if a model needed to know the structure of a protein it could use AlphaFold for that.

We think the combination of Gemini’s world models, AlphaGo’s search and planning techniques, and specialized AI tool use will prove to be critical for AGI.

True creativity is a key capability that such an AGI system would need to exhibit. Move 37 was a glimpse of AI’s potential to think outside the box, but true original invention will require something more. It would need to not only come up with a novel Go strategy, as AlphaGo impressively did, but actually invent a game as deep and elegant, and as worthy of study as Go.

Ten years after AlphaGo’s legendary victory, our ultimate goal is on the horizon. The creative spark first seen in Move 37 catalyzed breakthroughs that are now converging to pave the path towards AGI - and usher in a new golden age of scientific discovery.

China attack on Taiwan triggers a depression

When China invades Taiwan, the consequences will be global. Taiwan is the undisputed epicenter of the world’s chip supply, producing over 90 percent of most advanced semiconductors: the high-performance chips that power today’s AI, GPUs, robotics. These are also the chips that power your phones, computers, cars and medical devices. If those factories are seized or destroyed, the global economy will crash overnight. Tens of trillions of dollars in losses, supply chains in chaos, the worst economic depression in a century.

AI-automated military deters war

Palmer Luckey, Entrepreneur, CEO Andruil, January 2025, The AI Arsenal That Could Stop World War III | Palmer Luckey | TED,

Now while we make dozens of different hardware products, our core system is a piece of software, an AI platform called Lattice, that lets us deploy millions of weapons without risking millions of lives. It also allows us to make updates to those weapons at the speed of code, ensuring we always stay one step ahead of emerging and reactive threats.

Another big difference is that we design hardware for mass production using existing infrastructure and industrial base. Unlike traditional contractors, we build, test and deploy our products in months, not years.

That approach has allowed us, in less than eight years, to build autonomous fighter jets for the United States Air Force, school bus-sized autonomous submarines for the Australian Navy, and augmented reality headsets that give every one of our superheroes superpowers, to name just a few. We also build counter-drone technology like Roadrunner here, which is a twin turbojet-powered, reusable counter-drone interceptor that we took from napkin sketch to real-world combat-validated capability in less than 24 months. And we did it using our own money.

Now coming from a guy who builds weapons for a living, what I’m about to say next might sound counterintuitive to you. At our core, we’re about fostering peace. We deter conflict by making sure our adversaries know they can’t compete. Putin invaded Ukraine because he believed that he could win. Countries only go to war when they disagree as to who the victor will be. That’s what deterrence is all about. Not saber-rattling. Making aggression so costly that adversaries don’t try in the first place. So how do we do that?

06:19

For centuries, military power was derived by size. More troops, more tanks, more firepower. But over the last few decades, the defense industry has spent far too long handcrafting exquisite, almost impossible-to-build weapons. Meanwhile, China has studied how we fight. And they’ve invested in the technologies and the mass that counter our specific strategies. Today, China has the world’s largest navy, with 232 times the shipbuilding capacity of the United States; the world’s largest coast guard; the world’s largest standing ground force; and the world’s largest missile arsenal, with production capacity growing every single day. We’ll never meet China’s numerical advantage through traditional means, nor should we try. What we need isn’t more of these same systems. We need fundamentally different capabilities. We need autonomous systems that can augment our existing manned fleets. We need intelligent platforms that can operate in contested environments where human-piloted systems simply cannot. We need weapons that can be produced at scale, deployed rapidly and updated continuously. Mass production matters.

In a conflict where our capacity is our greatest vulnerability, what we really need is a production model that mirrors the best of our commercial sector: fast, scalable and resilient. We know how to win like this. We rallied our industrial base during World War II to mass produce weapons at an unprecedented scale. It’s how we won. The Ford Motor Company, for example, produced one B-24 bomber every 63 minutes.

But to actually achieve the benefits of these mass-produced systems, we need them to be smarter. This is where AI comes in. AI is the only possible way we can keep up with China’s numerical advantage. We don’t want to throw millions of people into the fight like they do. We can’t do it, and we shouldn’t do it. AI software allows us to build a different kind of force, one that isn’t limited by cost or complexity or population or manpower, but instead by adaptability, scale and speed of manufacturing.

Now the ethical implications of AI in warfare are serious. But here’s the truth. If the United States doesn’t lead in this space, authoritarian regimes will. And they won’t be concerned with our ethical norms. AI enhances decisio-making. It increases precision. It reduces collateral damage. Hopefully, it can eliminate some conflicts altogether. The good news is that the US and our allies have the technology, human capital and expertise to mass-produce these new kinds of autonomous systems and launch a new golden age of defense production.

With all that information in mind, let’s go back to Taiwan. But imagine a different scenario. The attack might begin the same way: Chinese missiles streak towards Taiwan. But this time, the response is instant. A fleet of AI-driven, autonomous drones, already stationed in the region by allies, launch within seconds. Swarming together in coordinated attacks, they intercept incoming Chinese bombers and cruise missiles before they ever reach Taiwan. In the Pacific, a distributed force of unmanned submarines, stealthy drone warships, and autonomous aircraft that work alongside manned systems strike from unpredictable locations. Our AI-piloted fighter swarms engage Chinese aircraft in dogfights, responding faster than any human possibly could. On the ground, robotic sentries and AI-assisted long range fires halt China’s amphibious assault before a single Chinese boot reaches Taiwan’s shores.

By deploying autonomous systems at scale, this type of autonomous system, we prove to our adversaries that we have the capacity to win. That is how we reclaim our deterrence. To do so, we just have to stand with our allies across the world, united by the shared values and common resolve that we’ve shared for the better part of a century. Our defenders, the men and the women who volunteer to risk their lives, deserve technology that makes them stronger, faster and safer. Anything less is a betrayal because that technology is available today. This is how we prevent a repeat of Pearl Harbor. We could be the second greatest generation by rethinking warfare altogether.

Autonomous warfare inevitable, humans can’t counter

Patraeus & Flanagan, March 12, 2026, David Petraeus is a Partner at the investment firm KKR and Kissinger Fellow at Yale University’s Jackson School. Between 2007 and 2011, he served in top U.S. military roles, including command of the surge in Iraq. Between 2011 and 2012, he was Director of the CIA. He is a co-author of Conflict: The Evolution of Warfare from 1945 to Gaza; Isaac C. Flanagan is Co-Founder of Zero Line, a nonprofit organization that works with international partners to identify critical needs in Ukraine’s defense sector, Foreign Affairs, The Autonomous Battlefield, https://www.foreignaffairs.com/middle-east/autonomous-battlefield

The era of autonomous warfare will not announce itself with robotic armies marching across battlefields. Instead, it is already emerging, quietly and inexorably, in the skies and fields of eastern Ukraine (and to a lesser degree in the Middle East), where missions are increasingly executed by machines at speeds no human can match and electronic warfare is severing the links between operators and their machines. Very soon, autonomous systems will no longer operate individually; over time, they will form platoon- or even battalion-sized units that share information and coordinate without human intervention. And the side that waits for human approval before acting will lose.

This transition demands that militaries rethink not just the nature of command but the fundamental nature of war. The adaptation challenge goes beyond technological and industrial issues, although those aspects are enormously important. Already, Ukrainian engineers are rapidly developing software for autonomous navigation, and Ukrainian military technicians are now assembling first-person-view drones and other types in extraordinary numbers: some 3.5 million last year and a potential seven million this year, compared with 300,000 to 400,000 now assembled annually in the United States. The U.S. military will have to adapt much faster to manufacture drones in the enormous numbers required and to learn to employ autonomous systems effectively.

But hardware and software will not be enough. It will be just as critical to develop new concepts and doctrine, adjust organizational structures, and institute the new kinds of military education and training that autonomous warfare will demand. These are all areas in which military institutions are often overly deliberate. But which militaries move first to change how they think about command and how the nature of war is evolving will determine which countries win the wars of the future.

MACHINE LEARNING

Unmanned systems in warfare exist on a spectrum, but not all of them are autonomous. At one end are remote-controlled systems: machines piloted or driven continuously by a human operator via a communications link. (Think of a Predator drone operator in Nevada piloting missions over Afghanistan.) Militaries began incorporating remote-controlled systems decades ago: unmanned target aircraft date back to World War I, and guided aerial weapons were operational by World War II. But the modern era of remote control began in 1995, when the Predator first flew reconnaissance missions over Bosnia. By 2015, the U.S. military was operating nearly 11,000 unmanned aerial vehicles, up from 90 in 2001; the Pentagon now plans to field more than 300,000. Today, an estimated 200,000 remote-controlled drones are being launched monthly in Ukraine, alongside unmanned surface vessels that have sunk Russian warships and, in one case, shot down fighter jets over the ocean.

But none of these systems, however impressive, are autonomous. They depend on a human at the controls. Autonomy begins when that human is no longer required—either because electronic warfare severs a system’s command-and-control link and onboard programming takes over or because the system no longer needs remote piloting to complete the mission. The autonomous threshold is already being crossed in Ukraine. Unmanned systems fielded by both Kyiv and Moscow increasingly default to onboard programming when jamming severs their communications links, continuing their missions until human control can be restored or the mission is complete.

Kyiv and Moscow alike have pushed the envelope on autonomy because electronic warfare and air defenses have become so pervasive in the operational environment. No commander can count on continuous human control. Ukrainian operators now routinely launch systems knowing that their control links will be jammed or spoofed within minutes. Their success depends on how well they have preprogrammed the onboard software that takes control when communications are cut. In a December 2024 Ukrainian assault on Russian forces near Kharkiv, Ukraine’s 13th National Guard brigade launched what was reported to be the first offensive operation conducted entirely with unmanned systems. Instead of deploying soldiers on the ground, remotely controlled ground vehicles advanced to lay and clear mines and fire on Russian defenses while surveillance, bomber, and suicide drones provided battlefield awareness and air support.

Unmanned systems fielded in Ukraine increasingly default to onboard programming.

The attack destroyed Russian defensive positions and ultimately enabled Ukrainian infantry to advance and seize ground they still hold today. Not a single soldier was exposed during the initial assault—and careful planning and disciplined communications meant that not a single autonomous system was lost to Russian jamming, either. This coordination was impressive. But it was still controlled by humans. Pilots based in separate locations watched shared video feeds and sequenced their actions manually, and the systems did not communicate with each other.

A much more fundamental shift is on the horizon: autonomy from launch. These systems will execute independently from the start of a mission. This is not the autonomy of a cruise missile or drone following a predetermined flight path to a fixed location. Autonomy from launch means systems that adapt their execution within commander-set constraints: coordinating with other elements in a formation, responding to changing conditions, and selecting among authorized actions when disconnected from human control, although humans will monitor their progress and retain the ability to retask or abort as long as communications remain open.

Currently, autonomy from launch exists only in fledgling form. Individual drones equipped with artificial-intelligence-assisted targeting—which can find and strike targets without an operator’s continuous control—number in the thousands among millions of remotely controlled systems. But over time, such machines will not operate as standalone units. Instead, commanders will mass them into formations—air, ground, and maritime systems that include drones, sensors, and targeting elements that direct and coordinate movement and strikes. These formations will execute the commander’s intent and preprogrammed directions even when disconnected.

MECHANICAL WAVE

Militaries worldwide now know that they need to produce many more drones. But they risk missing the deeper point. The advantage in the coming era will not go to the side that assembles the largest fleet of unmanned systems. It will go to the side that first develops the operational concepts to employ them—and then redesigns command-and-control systems, organizations, and training and operations to match. The technology is arriving. But the big ideas, the concepts, must arrive first.

The autonomous formation—whether it is the equivalent of a platoon-sized or battalion-strength fleet of autonomous systems—will integrate air, ground, and maritime systems with sensors, weapons, mobility, and protection. Not only will such formations be able to execute a commander’s intent at a remove and potentially out of contact, but they will also coordinate with each other at machine speed. This will radically change the traditional timing of battles and enable militaries to identify and exploit fleeting tactical windows faster than adversaries can respond—even if those adversaries have also deployed remotely piloted systems.

Consider the advantages: a military that possesses such synchronized systems—and deploys them carefully and effectively—can compress the time it traditionally takes for staff to detail strike options to commanders, for commanders to deliberate and then issue orders to subordinates, and for subordinates to relay orders to the pilots or drivers that are remotely controlling systems. In conventional high-intensity conflicts—the kind currently being waged in Ukraine—autonomous formations will be able to maintain offensive momentum even when electronic jamming severs communication links.

The sides that win future wars will not be the ones with the most drones.

Autonomous formations will transform irregular warfare (such as counterinsurgency campaigns in the Sahel or Gaza), stability operations (maintaining order in a postconflict environment), and gray-zone competition (such as the maritime pressure Beijing exerts in the South China Sea), too. In these scenarios, autonomous capabilities will enable far more persistent intelligence, surveillance, and reconnaissance by, for instance, allowing for the monitoring of vast stretches of territory around the clock as sensors automatically detect unusual movements or changes in patterns—tasks that today require rotating shifts of human analysts watching video feeds. Autonomous systems will also enhance force protection and precision strikes by continuously guarding personnel and collapsing the time between identifying an enemy target and striking it.

But especially in high-intensity conventional fights, such as the kind being waged in Ukraine, the compressed decision cycle that autonomy offers will transform how commanders orchestrate operations. Delegation and degraded-communications planning will become increasingly important. Such formations operating inside preset boundaries will maintain offensive momentum even when electronic warfare cuts links across entire sectors. The side that masters this and has enough unmanned systems will win. Any military that tries to retain human control of the tempo of battle, meanwhile, will experience a serious liability. Militaries will have to decide in advance which choices must remain under human control and which can be delegated to machines—and ensure that autonomous execution aligns with the commander’s intent when actions are faster than individuals can intervene. Ultimately, the winner will not be the side with the most drones but the side that best solves the command-design problem (and still has plenty of unmanned systems). To be sure, humans should retain certain key judgments, including when to escalate, how to engage populations, and whether a strike serves or undermines political objectives. In democracies, in particular, these decisions will need to remain irreducibly human across every type of conflict.

Data centers rely on fossil fuels and cause massive pollution

Wong, March 13, 2026, Matteo Wong is a staff writer at The Atlantic, Inside the Dirty, Disutpian World of Data Centers, https://www.theatlantic.com/magazine/2026/04/ai-data-centers-energy-demands/686064/

Already, the air smelled of soot, gasoline, and asphalt. Then I felt a tickle sliding up my nostrils and down into my throat, like I was getting a cold. As we approached, I heard the rumble of cranes and trucks, and then from behind a patch of trees emerged a forest of electrical towers. Finally, I saw it—a white-walled hangar, bigger than a dozen football fields, where Elon Musk intends to build a god.

This is Colossus: a data center that Musk’s artificial-intelligence company, xAI, is using as a training ground for Grok, one of the world’s most advanced generative-AI models. Training these models takes a staggering amount of energy; if run at full strength for a year, Colossus would use as much electricity as 200,000 American homes. When fully operational, Musk has written on X, this facility and two other xAI data centers nearby will require nearly two gigawatts of power. Annually, those facilities could consume roughly twice as much electricity as the city of Seattle.

To get Colossus up and running fast, xAI built its own power plant, setting up as many as 35 natural-gas turbines—railcar-size engines that can be major sources of smog—according to imagery obtained by the Southern Environmental Law Center. Pearson coughed as we drove by the facility. The scratch in my throat worsened, and I rolled up my window.

xAI’s rivals are all building similarly large data centers to develop their most powerful generative-AI models; a metropolis’s worth of electricity will surge through facilities that occupy a few city blocks. These companies have primarily made their chatbots “smarter” not by writing niftier code but by making them bigger: ramming more data through more powerful computer chips that use more electricity. OpenAI has announced plans for facilities requiring more than 30 gigawatts of power in total—more than the largest recorded demand for all of New England. Since ChatGPT’s launch, in November 2022, the capital expenditures of Amazon, Microsoft, Meta, and Google have exceeded $600 billion, and much of that spending has gone toward data centers—more, even after adjusting for inflation, than the government spent to build the entire interstate-highway system. “These are the largest single points of consumption of electricity in history,” Jesse Jenkins, a climate modeler at Princeton, told me.

Even conservative analyses forecast that the tech industry will drop the equivalent of roughly 40 Seattles onto America’s grid within a decade; aggressive scenarios predict more than 60 in half that time. According to Siddharth Singh, an energy-investment analyst at the International Energy Agency, by 2030, U.S. data centers will consume more electricity than all of the country’s heavy industries—more than the cement, steel, chemical, car, and other industrial facilities put together. Roughly half of that demand will come from data centers equipped for the particular needs of generative AI—programs, such as ChatGPT, that can produce text and images, solve complex math problems, and perhaps one day inform scientific discoveries.

To power AI, energy and tech companies are turning to fossil fuels, which they regard as more reliable and readily available than wind, solar, or nuclear. Asked where the energy for data centers should come from, OpenAI CEO Sam Altman has repeatedly said, “Short-term: natural gas.” (OpenAI and The Atlantic have a corporate partnership.) A Louisiana utility plans to build three natural-gas plants for a Meta data center that, upon completion, will be among the largest in this hemisphere. The lifespans of coal plants, too, are being extended to power new data centers. And the IEA estimates that data-center emissions could more than double by 2030—becoming one of the fastest-growing sources of greenhouse gases in the world.

The optimist’s case is that, by then, advanced nuclear reactors will have obviated many of the new fossil-fuel plants, and AI tools will have invented technologies that can solve the climate crisis. That may well happen. But today, “the market has converged on Add gas now, and then add nuclear later,” Jenkins said. In other words, if natural-gas turbines seem to offer the most expedient path to an AI-enhanced future, then clean air may have to wait.

Companies will have no control over how AI is used

Mark Anderson, March 11, 2026, Ross Andersen is a staff writer at The Atlantic. He was previously the magazine’s deputy editor. As a writer for the magazine, he has reported from Greenland, Russia, India, Pakistan, China, South Korea, and Japan. He is also the author of The Long Search, forthcoming from Random House, The Atlantic, Dario Amodei’s Oppenheimer Moment, https://www.theatlantic.com/technology/2026/03/anthropic-dod-ai-utopianism/686327/

Amodei does not explain precisely how the AIs will accomplish this. In most cases, he expects them to do what the smartest humans do, but much more rapidly, compressing decades of scientific progress. He says that by 2035, we could have the theories, cures, and technologies of the early 22nd century. Our infectious diseases and cancers could be cured, and we could live twice as long, and slow the decay of our brains. Demis Hassabis, the head of Google DeepMind, has similarly conceived of superintelligent AI as the ultimate tool to accelerate scientific discovery, and Sam Altman, OpenAI’s CEO, has said that advanced AI may even solve physics.

Amodei does not say that this utopian AI future is inevitable. To the contrary, among the chief executives at the top AI labs, he may be the one who worries most about the technology’s dangers. “Machines of Loving Grace” is an optimistic outlier in his larger oeuvre of published writing, much of which concerns the risks that will accompany the creation of a greater-than-human intelligence. Amodei seems to think of today’s AI researchers as comparable to Manhattan Project scientists, and has been known to recommend The Making of the Atomic Bomb. In his telling, superhuman AI could be even more dangerous than nuclear weapons, which is why AI needs to be developed the right way, by the right people, so that it doesn’t overpower humanity or tip the global balance of power toward autocracies.

Implicit in this vision is the hope that in the end, when the chips are down, Amodei, or someone very much like him, will have some say in how AI will be used. But if Anthropic’s recent experience with the Pentagon is any indication, that likely won’t be his decision to make. For all of Amodei’s reading and thinking about the early nuclear age, he may not have fully internalized its meaning.

…

Edward Teller, the father of the hydrogen bomb, also imagined a great many uses for atomic energy, including the violent reshaping of the Earth’s surface. …

The men who split the atom were right to believe that they were delivering humanity into a new world. …

Teller did get to live out his dream of sculpting land forms with nuclear explosions. Project Plowshare, the Atomic Energy Commission’s program devoted to these peaceful detonations, conducted 27 separate tests, but it achieved nothing except contamination and the galvanization of the environmental movement. Meanwhile, nuclear weapons did not shock the world’s leaders into a new era of peace and candor, as Bohr had hoped. A year before the bombing of Hiroshima, Bohr went to the White House to make his case for international openness on these matters to Franklin D. Roosevelt. The president, who was in his last months, was by some accounts sympathetic to Bohr’s arguments, but when Winston Churchill heard them, he was horrified. Churchill told an aide that Bohr should be locked up for even suggesting that the allies reveal their nuclear hand.

After the war, J. Robert Oppenheimer revived some of Bohr’s ideas and channeled them into a proposal for a new international agency that would control all dangerous nuclear activities. A similar plan was presented to the United Nations in June 1946. The Soviet Union rejected it and countered with a proposal that America simply destroy its arsenal first.

The United States did not destroy its arsenal but rather grew it, and developed new weapons that are more than 1,000 times more powerful than the one that leveled Hiroshima. Today, nine nations possess nuclear arsenals, comprising more than 12,000 warheads in total, including many that are set on a virtual hair trigger. The constant possibility that these arsenals will be used in a major exchange is the true lasting legacy of the nuclear age. The final remaining treaty constraining the two largest of them, belonging to America and Russia, expired last month without being replaced. Like the proposals put forth by Bohr and Oppenheimer, the treaties were defeated by the cold logic of competitive advantage, which will also likely shape the global future of AI.

On February 27, Amodei released another piece of writing, a memo for a smaller audience: his staff. Four days earlier, the Pentagon had issued an ultimatum demanding that Anthropic remove any restrictions on how the military used its AI model, beyond existing law. The model had been operating on America’s classified networks since last year, and reportedly has already been used in America’s attacks on Venezuela and Iran.

It’s striking that only a few years into the large-language-model moment, these models seem to have become central to the most complex operations of the world’s most powerful military, but Amodei has no general objection to AI’s use in war. He had eagerly sought a Department of Defense contract, in part because he believes that democracies should use AI to maintain a military edge over China and the world’s other autocracies, which will almost certainly be using AI more and more in the years to come.

Amodei had stuck to two red lines throughout his negotiations with the Pentagon: He didn’t want the awesome informational processing power of Anthropic’s AI used for mass surveillance of American citizens, and he didn’t want it directing autonomous weapons that could kill without human oversight. The Pentagon refused, demanding unrestricted use of Anthropic’s model, Claude. After the talks broke down, it used a coercive tool never before deployed against an American company, a supply-chain-risk designation, which could imperil Anthropic’s business. (Anthropic has since filed suit to have it removed. The company declined to comment for this article.) And while all of this was happening, Altman swooped in to finalize his own Pentagon deal for OpenAI.

Amodei’s frustration with the week’s events leaked into the memo that he wrote to his staff. Its tone differed greatly from “Machines of Loving Grace.” Amodei excoriated OpenAI, and described the reported provisions of its deal as “safety theater.” (OpenAI later added what it has said are stiffer provisions to its deal.) The haste with which OpenAI’s leadership had come to an agreement with the Pentagon clearly irked Amodei; the episode revealed “who they really are,” he said.

But Amodei didn’t seem to reckon with the larger structural lesson here. Anthropic’s dispute with the Pentagon is a reminder that the people who create a powerful technology don’t usually get the final say in how it’s used. The models aren’t even all that advanced compared with what they will be, and in Venezuela and Iran, the U.S. is not facing off against the world’s great AI champions. Yet the Pentagon still bristled at the very idea that its use of Anthropic’s AI could be limited, and in the face of resistance, it threatened to burn a private company down. If AI becomes a much more dangerous weapon, and the U.S. finds itself pitted against a country with frontier models that are as powerful as its own, the government will almost certainly demand total control or commandeer the technology outright.

After the builders of the atomic bomb finished their work in the New Mexico desert, they very quickly learned how little say they would have in its use. The weapons were driven away on trucks, and in the weeks afterward, no one called the scientists to get a green light for the bombing of Hiroshima or Nagasaki. Neither did anyone ask them to sign off on future additions to America’s nuclear stockpile. Their leverage was front-loaded: They could choose to create their terrible weapons or not, but once they’d successfully tested even one, they’d already forfeited it.

Amodei now finds himself in a similar position. He may well be right that soon, whole “countries of geniuses” will occupy the data centers that are being built, en masse, all over the world. But whether anyone will be able to control such a technological force remains an open question, and either way, it certainly won’t be him.