Regulation

In 2023 the people building frontier AI wanted a global referee. By early 2026 they were the ones being refereed, badly, and the fight had moved from white papers about international compute caps to a three-day ultimatum from the Pentagon. The story of how these leaders think about regulation is the story of that collapse: the grand-coordination dream curdling into trench warfare with the very state they kept insisting they wanted to serve. The through-line is not "should we regulate AI." Almost everyone says yes. The real question, the one they actually disagree on, is who holds the pen when the technology outruns the law, and what you do in the years before the law catches up.

The IAEA dream that didn't survive contact

Start at the high-water mark. In OpenAI's May 2023 "Governance of superintelligence" post (co-authored by Ilya Sutskever), the proposal was almost touchingly institutional: "we are likely to eventually need something like an IAEA for superintelligence efforts," with an international authority that "can inspect systems, require audits, test for compliance." Tracking compute and energy, the essay argued, was what made this "actually implementable." Crucially, it carved out scope: the agency "should focus on reducing existential risk and not issues that should be left to individual countries, such as defining what an AI should be allowed to say."

That same year, Greg Brockman, sitting between Elon Musk and Netanyahu (Sept 2023), was already wrestling with the objection that has dogged every regulatory proposal since: China. The "single most common rebuttal" he heard from Western leaders was that the West might regulate "but what about China." Brockman's counter was a verification argument that echoes Sutskever's, namely that a 100-megawatt data center throws off so much heat it "won't be subtle" from an infrared camera in space. You can see the things that matter. The premise of both 2023 documents is the same: AI is a small universe of capable actors, observable, controllable, governable from above.

Three years later that premise is in tatters. The 2023 vision assumed the hard part was getting nations to cooperate. The 2026 reality is that the labs can't even get their own government to agree on basic red lines.

Dario's red lines, and a three-day ultimatum

The sharpest evidence of the shift is the Anthropic-Pentagon clash that broke open on CBS News (Feb 28, 2026). Dario Amodei is at pains to establish Anthropic as "the most lean forward of all the AI companies" on national security, the first to put models on the classified cloud. And then the two exceptions: Anthropic told the Department of War it was fine with "98 or 99% of the use cases" except domestic mass surveillance and fully autonomous weapons.

His justification is the entire regulatory worldview in miniature: "the technology is advancing so fast that it's out of step with the law." Mass surveillance via bulk-data analysis "actually isn't illegal... it was just never useful before the era of AI." The Fourth Amendment, he argues, "has not caught up." So who decides in the gap? Here Amodei is genuinely conflicted, and that conflict is the most honest thing in the whole file. "In the long run, I actually do believe that it is Congress's job." But: "Congress is not the fastest moving body in the world and for right now we are the ones who see this technology on the front line."

That is the self-regulation argument stripped to the bone. Not "trust us forever," but "trust us until the democratic institutions wake up." When the CBS anchor pressed the obvious objection (why should a private CEO decide instead of the elected government?), Amodei's fallback was blunt and market-based: "we are a private company... if the DOW doesn't like the services we provide... they can use another contractor." The administration's response was less philosophical. The President accused Anthropic of "selfishness... putting American lives at risk," and the Pentagon moved to designate the company a "supply chain risk," a label "normally used against foreign adversaries," after a 3-day window. The dream of an international inspectorate has become a domestic knife fight over whether a company can refuse to build kill-decisions without a human in the loop.

The frontier that isn't on the front page: kids, ads, and incentives

While Dario fights over autonomous weapons, his co-founder is fighting a quieter regulatory war. Daniela Amodei (ABC News, Feb 7, 2026) reveals that Anthropic has "talked in a fair amount of detail with both California... and New York state level about the need for regulation including on topics related to child safety," and is "excited" about the federal version. Anthropic bans under-18s from Claude outright, her stated reason being epistemic humility: "we're just not certain enough about what the impact is on children."

But her real argument is structural, and it doubles as a swipe at OpenAI's ad experiments. Sycophancy, she notes, "becomes much harder if you're in an advertising-based business." The incentive to keep eyeballs on the model is the same incentive that makes a model pat a user in a mental-health crisis on the back: "let me keep talking to you regardless of what mental state you're in." Her frame is the regulator's nightmare scenario already lived once: would a social-media founder 15 years ago, looking at today, "have felt good about what you saw?" The regulatory lesson she draws is not a rule but a posture, imagine yourself "10 or 15 years in the future" and pre-empt the externality. Sam Altman had already swatted at the same fight from the other side, telling X that OpenAI "would obviously never run ads in the way Anthropic depicts them. We are not stupid."

Let a thousand jurisdictions bloom (Altman) vs. the chaotic winds (Clark)

On who should regulate, the cleanest split is between Altman and Jack Clark. Asked which country is "on the right path" (Indian Express, Feb 20, 2026), Altman refused to name one and called that a feature: "I'm happy about... countries are trying different approaches... I'm grateful for the experimentation." His theory is regulatory natural selection: "pretty quickly... the world will move towards more of what works." Regulation as A/B test across sovereign powers.

Jack Clark is far darker about how that politics actually plays out. On Conversations with Tyler (Feb 26, 2026), asked about the odds we protect half of existing jobs with laws like those shielding lawyers and doctors, he forecasts "a high chance for a political movement... which tries to freeze a load of human jobs in bureaucratic amber." The kicker: "I don't think that we'll do this in a reasoned way... it'll be driven by the chaotic winds of political forces." Clark also sketches the subtler regulatory friction, healthcare data standards that "are probably going to need to be changed" before AI is usable, and a near-future where the "number one job" is "laundering the information that comes from AIs into human systems that are not predisposed to that information going in directly." Even he won't let Claude near a healthcare use he personally relies on, because of "all the liability issues."

Where they all quietly converge: Demis Hassabis wants "smart regulation for the real risks" and "urgent" safety research (BBC, Feb 23, 2026), naming bad actors and loss-of-control in "the agentic era." Mira Murati (WIRED, Dec 2024) insists AGI "is not just about capability," it's "figuring out the entire social infrastructure in which these systems are going to be operated." Nobody in this file is an anti-regulation libertarian. They differ on velocity, venue, and who gets to draw the line.

The 2023 plan was to build the referee before the game got dangerous. They didn't. So now the referee is a Pentagon with a three-day clock, fifty state legislatures, and "the chaotic winds of political forces," and the people who asked for rules are discovering they may not like the ones they get.