OpenAI and Anthropic Push for Global AI Regulation at the UN as Washington and Beijing Resist

Sam Altman of OpenAI and Dario Amodei of Anthropic brought their case for binding global AI regulation to the United Nations Security Council during this year's General Assembly in New York. Both told member states that pledges developers make on their own have reached the limit of what they can deliver, and that verifiable international standards are needed before frontier systems outrun human oversight. The request puts the two labs in direct conflict with the governments that host them.

Leadership from Hugging Face joined the delegation, which extended the appeal beyond the two largest US labs. The message to the council was the same from all three: development is moving faster than any single government can track, and the risks to societies require coordination no country can supply alone. Altman and Amodei described the moment as a turning point and warned that fragmented oversight leaves AI harder to control.

Washington answered quickly. Michael Kratsios, a senior technology adviser in the Trump administration, rejected the proposal, arguing that new global governance structures are unnecessary and that testing and review of frontier models belong inside US agencies. Beijing took a similar position and declined to endorse additional international architecture. Those responses carry outsized weight because the two countries with the most advanced AI sectors both hold permanent Security Council seats, and the council is the only UN body whose decisions can be made binding.

The deadlock has a structural cause. The governments able to make rules stick are also the ones whose domestic industries would carry them, so any council text that binds the labs binds the competitive assets of their hosts at the same time. The companies asking for rules sit closest to the frontier, and they are also the best positioned to pay for them.

The pitch for global AI regulation

What the labs want is narrower than a treaty and harder to dismiss than a policy paper. OpenAI and Anthropic are asking for shared, verifiable standards covering the most capable models, with testing and evaluation that outside governments can inspect instead of accepting on trust. Both companies already publish safety frameworks. Those documents are written by the companies themselves, can be revised at any time without outside consent, and carry no force beyond the terms of service attached to their products.

Three signatures carry less weight than a dozen. The appeal came from OpenAI, Anthropic and Hugging Face, leaving other large model developers outside the room. A standard endorsed by the two most prominent US labs and a widely used model repository is a starting point rather than a consensus, and governments can reasonably ask why the companies closest to the frontier are the ones requesting common rules.

Voluntary commitments have a structural weakness the executives themselves pointed to: they bind only the firms that sign them and weaken the moment a competitor decides restraint costs more than it protects. An international standard changes that arithmetic by turning restraint into a market-wide condition instead of a unilateral handicap.

Washington and Beijing push back

Both sides agree the technology is dangerous. The disagreement is over where authority should sit. The US position rests on a specific claim: domestic agencies can evaluate frontier models faster and with more technical depth than any multilateral process. That claim has force where classified information, export controls and chip supply chains are involved. It holds up less well at the point of deployment, since a model trained under one jurisdiction makes decisions in dozens of others.

What Washington offers instead is oversight located somewhere else. Agency-level evaluation, export controls and procurement conditions already shape what frontier developers can build and ship. Those instruments are unilateral, can be changed by executive action, and stay largely invisible to other governments until they are applied, which is the objection the labs raised.

The venue compounds the difficulty. A Security Council resolution is the shortest route to binding language, but it requires the concurrence or abstention of all five permanent members. With Washington and Beijing both opposed to new structures, the realistic near-term output is a statement, a working group, or a request for further study. An enforceable regime is not on the table.

The verification problem

Compliance is where the debate becomes concrete. Verifiable standards imply compute thresholds, evaluation reports, incident disclosure, and third-party auditors with access to model weights or inference logs. For frontier labs with large safety and legal teams, that is overhead. For smaller developers and open-weight projects, the same requirements can be prohibitive, which is why binding global AI regulation is read by its critics as a competitive moat as much as a safety measure.

Who performs the verification matters as much as what gets verified. A regime built on national agencies inherits the enforcement gaps of the countries that host those agencies. A system anchored in a new UN body inherits a mandate and a budget that member states have not agreed to fund. Neither option resolves the underlying tension, because AI oversight is now treated as defense-adjacent policy, and defense-adjacent policy rarely moves into open multilateral verification.

There is a cost asymmetry that council members will notice. Rules that bite hardest on frontier training runs raise the entry price for everyone below that threshold, and they fall most heavily on developers who lack the capital to run compliance programs at scale. That does not make the standards wrong, but it explains why the request from two well-capitalized labs lands differently in capitals with younger AI industries.

Incumbency also shapes the request. Firms already operating at the frontier have less to lose from rules that raise the cost of reaching it, and more to gain from a framework that makes their own safety investments legible to enterprise buyers and regulators. That does not discredit the safety argument, but it explains why support for binding global AI regulation is strongest among the companies best positioned to absorb it.

Open-weight releases sharpen the problem further. Once weights are public, no verification regime can withdraw them, and the only remaining lever is the supply chain that produced them.

What changes from here

Pressure for action is building outside the industry. The Australian government has attributed a website breach to an agent built on OpenAI's technology, giving governments a concrete case to cite when they argue that current oversight is inadequate. Daniel Forti, who leads UN affairs at the International Crisis Group, has argued that member states increasingly recognize a need for international cooperation and rule-setting on AI, even where the largest players remain more focused on growth than on collaboration.

Enterprises in regulated industries sit on the receiving end of that uncertainty. Buyers assembling AI systems for finance, health, or public services need to document which evaluation standard a model was tested against, and conflicting national requirements mean duplicate testing for the same deployment.

The next test is procedural rather than philosophical. If the Security Council moves from testimony to a formal text, the question becomes which obligations survive negotiation and which are softened into reporting requirements. If it does not, the center of gravity shifts back to national regulators, and the industry's own request becomes another entry in the record of proposals the two dominant AI powers declined to adopt.

Why this matters

The two labs have inverted the usual argument about AI oversight: the objection that rules would slow innovation now has to contend with the builders asking for them. What remains unresolved is whether the Security Council, whose permanent members include the two governments most resistant to new AI governance, can produce anything binding. For technology leaders, the practical implication is that compliance planning should assume a patchwork of national regimes for the next several years, with international standards arriving, if at all, as a floor rather than a ceiling.

Photo by Salya T on Unsplash

Related Articles

• OpenAI's Reversal: The Push for Mandatory National AI Safety Rules
• Anthropic Pushes for a Mandatory AI Kill Switch While the White House Dismisses Safety Fears
• OpenAI national security principles define red lines for AI weapons

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.