Altman and Amodei Asked the UN for Global AI Oversight. Washington Said No.

We build agentic products on top of models from both labs at the center of this story, so the question of who gets to set the rules those labs operate under isn’t a policy abstraction for us — it’s the regulatory environment our own roadmap has to plan around. On September 23, the UN Security Council held a session on artificial intelligence during the 81st General Assembly in New York, convened by France, and it put the leaders of the two most valuable AI labs in the same room as national governments to argue, in public, about who should decide how AI gets built and deployed, per Al Jazeera and CNBC.

What was said

• Sam Altman (OpenAI) warned that humanity risks “losing control of the future of AI” and argued the decisions at stake can’t be left to a handful of San Francisco labs — he called for governance rooted in “democratic processes and… governments accountable to the people.”
• Dario Amodei (Anthropic) went further on risk: “If managed poorly, I even believe AI could be a risk to humanity as a whole.”
• Yoshua Bengio, the Turing Award-winning researcher who’s become one of the field’s most prominent safety voices, called it an “unprecedented threat.”
• UN Secretary-General António Guterres framed the stakes against concentrated private power directly: decisions about AI “cannot be decided by a handful of countries or left to the whims of a few billionaires,” per ABC News.
• Clement Delangue (Hugging Face) described actively defending against AI-driven attacks built on Chinese open-weight models — a detail that ties this story directly to the broader US-China AI competition playing out this week.

Washington’s answer came from Michael Kratsios, Director of the White House Office of Science and Technology Policy, representing the US at the session: “We totally reject all efforts by international bodies to assert centralised control and global governance of AI.” Coverage framed the exchange as Trump rebuffing what his administration calls a “globalist scheme” to control AI development, per the Washington Post.

The reversal worth noticing

Industry lobbying against regulation is the default pattern. This is the opposite: the CEOs actually building frontier AI are now publicly to the left of the US government on international oversight, asking for exactly the kind of regime Washington is refusing to accept. The administration’s position isn’t new — since taking office it has prioritized “winning” the AI race against China through deregulation and infrastructure buildout over binding multilateral frameworks — but having it stated this bluntly, in direct response to the labs’ own CEOs, is new.

The tension we’re not going to paper over

There’s a real skepticism case against the CEOs here, and we’re not going to pretend it doesn’t exist: if frontier AI is genuinely as dangerous as Amodei and Bengio describe, that’s also an argument for why only a small number of well-capitalized incumbents should be trusted to build it — a moat dressed as a safety position. Delangue’s comment about defending against Chinese open-weight models sits awkwardly next to that, since “global oversight” proposals tend to target open-weight distribution specifically, and Hugging Face’s business depends on that ecosystem staying open.

There’s also a real case for Kratsios’s rejection that isn’t just reflexive anti-multilateralism: AI capability is software, not enriched uranium — it diffuses fast and is hard to monitor at any border, and nuclear non-proliferation analogies get invoked constantly without anyone specifying a workable enforcement mechanism. The US bet is that it retains more leverage moving fast and unilaterally, paired with chip export controls, than it would ceding any ground to a UN-brokered framework China would also sit inside. Both of those arguments can be true at once, which is exactly what makes this an open fight rather than a settled one.

What we’re watching

Whether the EU or UK use this session as a launchpad for an actual proposal, or whether it fades as a one-off statement. Whether Anthropic or OpenAI change any real product or deployment decisions in ways that reflect the risk posture their CEOs just described in public — that’s the credibility test for the rhetoric, not the speech itself. And whether the informal Bessent–He Lifeng US-China “AI hotline,” a bilateral channel separate from the UN process entirely, ends up being the more operationally real venue for AI-safety coordination between the two AI superpowers — a question that gets a data point today, with Trump and Xi meeting in Washington. For us, the practical stake is simple: whether the labs we build on top of end up operating under one enforceable global standard, or a patchwork of national rules with real jurisdictional gaps between them, is going to shape what “compliant” means for everything we build downstream of their APIs.