OpenAI and Anthropic ask the UN for global AI rules. Washington says no.

Two chief executives who compete for the same customers, the same researchers and the same headlines stood in front of the UN Security Council on 23 September 2026 and said, in near-identical terms, that their industry cannot be trusted to govern itself. Sam Altman flew to New York and spoke in person; Dario Amodei appeared by video link; Hugging Face's Clément Delangue joined remotely. France, holding the Council presidency for the month, had convened the session specifically to discuss AI and international security, and French Foreign Minister Jean-Noël Barrot chaired it personally. The unusual part was not that AI executives asked for regulation — labs have done that performatively for years — but the specificity of what they asked for, and the flat rejection waiting for them a day earlier from the country where both companies are headquartered.

Amodei's remarks were the starker of the two. He told the Council that unmanaged AI development was not an abstract worry but the defining security question of the decade.

He then set out three concrete mechanisms rather than a general appeal. First, narrow global agreements on specific catastrophic uses — a ban on using AI to design or manufacture biological weapons was his named example, chosen because it is a use nobody defends and every government has an incentive to prevent. Second, evaluation and verification systems that let countries check each other's stated safety commitments, rather than taking a lab's or a government's word for it. Third, common global standards for testing frontier models, paired with a notification system so a serious AI security incident in one country reaches the others before it repeats itself. Read together, that is a request for an inspection regime, not a treaty banning AI outright — closer in spirit to arms-control verification than to a moratorium.

Altman's framing leaned more procedural than existential, though he did not shy from the stakes. His core ask was for international standards covering four things at once: measuring what a model can actually do, assessing the risks that follow from that, judging whether a lab's safeguards are adequate, and preserving human oversight as systems get more capable.

He paired that with a practical ask for accurate, fast incident reporting — so, in his words, the world can learn from failures before they become catastrophes — and secure channels between governments, infrastructure operators and technical experts to share emerging threats. Altman also drew a pointed distinction that Amodei did not: any global framework, he argued, has to run through democratic institutions accountable to their own populations, which quietly excludes governments racing on AI without that accountability. He also made a point of not overclaiming. Any chance of an AI-caused catastrophe was unacceptable, he said, but he did not want to fall into the 'trap of doomerism' or the 'trap of blind optimism' — the two poles the AI safety debate has split into this year.

Here is the proposal set side by side, because the differences matter as much as the overlap:

Altman and Amodei were not speaking into a vacuum; they were speaking after the loudest possible answer to their question had already been given. On 22 September, addressing the General Assembly, President Trump rejected the entire premise of international AI oversight and announced that the US government would rename artificial intelligence 'super intelligence' in its own documents.

He was equally direct about why: growth, not caution, was the priority. 'I'm not going to stifle growth of something that will be bigger than the industrial revolution,' he told the chamber. The next afternoon, at the Security Council itself, the US delegate — Michael Kratsios, director of the White House Office of Science and Technology Policy — held that same line in the room where Altman and Amodei had just made their case, telling members the US would 'totally reject all efforts by international bodies to assert centralised control and global governance of AI.' No other delegation used the term 'superintelligence'; every other speaker, including the two CEOs, said 'AI'.

Strip away the diplomacy and the three ideas on the table — narrow bans, mutual verification, shared testing standards with incident reporting — describe something with real precedent: an inspection-and-disclosure regime, the model used for nuclear non-proliferation and, more recently, for the state-level AI rules starting to appear inside the US itself. Barrot's comparison to 1945 was not just rhetoric; the mechanisms he, Amodei and Altman were all gesturing at — independent verification, standardised testing, mandatory incident reporting — are the same three legs the nuclear order eventually stood on, adapted to a technology nobody can physically inspect a warehouse for.

The gap between the ask and the room is the story. Amodei and Altman's proposals only work if the countries hosting frontier labs sign on, and the country hosting the most consequential ones just told the room, twice in 24 hours, that it will not. China's representative, for its part, criticised what it called a 'petty us-against-them narrative' aimed at preserving a technological monopoly — a real disagreement about who global rules would actually constrain, not a shared objection to the idea of rules. Chinese developers DeepSeek and Moonshot were invited to address the Council; it is unclear whether either did. An UN framework with the US and China both unwilling to bind themselves is, in practice, a framework that governs the labs least likely to cause the incidents everyone is worried about.

The domestic US picture split the same day. Hours after the Security Council session, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act — legislation with teeth Altman and Amodei's UN remarks pointedly lacked, including a temporary development pause and prison terms for violations. It has no realistic path through this Congress, but it demonstrates that the appetite for binding rules the CEOs invoked abstractly at the UN already exists, concretely, in domestic politics — just not currently in the executive branch. And one further data point belongs in the same frame: Politico reported on 24 September that the White House had asked both OpenAI and Anthropic to hold new frontier models back from the UK's AI Security Institute pending a US review, with Anthropic already complying. A government that rejects multilateral AI oversight in public is simultaneously tightening bilateral control over who else gets to see the models first.

What to watch: whether France or another Council member tries to turn Wednesday's briefing into an actual draft resolution — which would require US and Chinese abstention at minimum to survive a veto, an unlikely combination given both governments' stated positions; whether Amodei's narrow-ban idea gets picked up bilaterally even without a UN vehicle, since a bioweapons-use ban is the one proposal with no obvious opponent; and whether the AISI access dispute widens into the first real test of whether 'international cooperation on AI safety' survives contact with a government that has just rejected the principle in the room built for exactly that purpose.