A first-of-its-kind meeting, an all-too-familiar split
On September 23, 2026, the United Nations Security Council did something it had never done before: it devoted an entire session to the risk that artificial intelligence could slip beyond human control. The witness list read like a roll call of the industry's most powerful voices. OpenAI's Sam Altman, Anthropic's Dario Amodei, Hugging Face's Clément Delangue, and Yoshua Bengio — the computer scientist who chairs the UN's Independent International Scientific Panel on AI — all addressed the world's most powerful diplomatic body. In the history of technology, there are few precedents for the people building a thing to warn, in formal session, that the thing might overwhelm its builders.
And yet, the substance of the session was eclipsed by the reaction to it. The day before, at the UN General Assembly, President Donald Trump had dismissed international AI oversight as a "globalist scheme," declaring he would not "stifle growth of something that will be bigger than the industrial revolution." At the Security Council itself, White House science adviser Michael Kratsios was blunter still: the United States, he said, "totally rejects all efforts by international bodies to assert centralised control" over AI. So the world's leading forum for collective security heard, in the same week, that AI could threaten humanity as a whole — and that the country hosting most of the companies building it wants no part in governing it together.
This is the defining contradiction of AI policy in 2026, and it deserves a clear-eyed assessment. The split is not, as some coverage implies, between people who take AI risk seriously and people who do not. The technology executives testifying before the Council clearly take it seriously. The split is between two incompatible theories of how to keep powerful AI safe: the race-to-the-front theory, which holds that American speed is itself the safety strategy because losing to China would be worse than any accident, and the coordination theory, which holds that no single country can make AI safe on its own because the technology respects no borders. Both camps cannot be right. The uncomfortable truth is that both camps may be half right — and that the policies currently on offer are built for neither.
The strange position of the AI chiefs
Consider the remarkable posture of the executives who testified. Amodei told the Council that AI "could be a risk to humanity as a whole" if managed poorly, and pledged that Anthropic would "slow down as much as necessary." Altman said OpenAI had "unilaterally slowed down in the past" and would do so again. These are extraordinary concessions from companies racing to ship the next generation of models. They are also, in practical terms, almost meaningless as policy. A unilateral promise to slow down is a promise with no enforcement mechanism, no timeline, and no penalty for reversal. It is, at best, a statement of intent — and at worst, a performance of responsibility staged for regulators.
The cynicism is not entirely unfair. Amodei had already laid out his position in a widely discussed essay earlier in September, in which he called for a coordinated global slowdown of AI development, overseen by third-party evaluators, to ensure "companies take adequate time to align and safeguard their models." The essay is ambitious: it asks the United States to ban the sale of powerful AI chips and semiconductor manufacturing equipment to China, crack down on chip smuggling and model distillation, restrict remote access to overseas data centers — and then sit down with Beijing to agree on pre-release testing of models for acute risks in cybersecurity, biology, and alignment, administered through a new global standards body. Even Amodei concedes the hardest part: giving such a body "real teeth will be a challenge, and the difficulty will be in verification that both sides don't have secret models which they don't test but may deploy in secret."
The critics of this vision have a point that its supporters should take seriously. As Adam Thierer of the R Street Institute has observed, there is a "realpolitik of global technological control that must be taken into account." Arms-control history is not encouraging. Chemical, biological, and nuclear regimes all wrestle with verification, and those deal with physical materials that can be inspected. AI models are mathematics; they can be copied onto a hard drive and hidden in a basement. A treaty that cannot detect cheating is not a treaty; it is a speech. And there is a second, sharper critique, voiced with characteristic bluntness by the libertarian press: if the executives are so worried, why do they not simply slow down themselves, rather than asking governments to coerce their competitors into slowing down? The suspicion — that the calls for regulation are, at least in part, calls to pull the ladder up behind the incumbents — is not proof of bad faith, but it is a reasonable question that deserves an answer rather than a dismissal.
The case for the American "no"
Now consider the administration's position on its own terms, because it is more coherent than its critics allow. The central claim is about China. Treasury Secretary Scott Bessent put it in apocalyptic terms at a Washington event this month: "Beating China — there is no day after tomorrow if China wins at this. If they were to pull ahead of us on AI, then nothing else matters." Strip away the rhetoric and the argument is this: AI is the decisive general-purpose technology of the century; the country that leads in AI will lead in economic output, military capability, and standard-setting; any international regime that slows American development while Beijing defects from it is not a safety measure but a forfeiture. From this vantage point, Kratsios's rejection of "centralised control" is not anti-safety but pro-sovereignty: the United States will manage AI risk through its own export controls, its own evaluations, and its own industrial policy — and will not subcontract that judgment to a body where its rivals hold vetoes.
There is also a defensive, institutional logic at work. International institutions have a poor record of governing fast-moving technologies, and a global AI standards body would inevitably become an arena for great-power contest — each side pushing standards that favor its own industry. The administration's bet is that voluntary pledges from American companies, combined with hard export controls on the chips and tooling that make frontier training possible, can keep the race won without slowing the runners. One can disagree with that bet. But it is not nothing.
The case against walking away
The problem is that the bet has three visible flaws. First, export controls are a speed bump, not a wall. Chip smuggling networks are already sophisticated, model weights can be stolen or distilled, and algorithmic progress means each generation of models requires less compute than the last. Amodei himself, notably, pairs his export-control asks with the admission that the real race is also a software race. Second, the "no international regime" position does not eliminate international governance; it merely surrenders the pen to others. China, the European Union, and a growing bloc of countries are writing their own rules, and American companies will have to comply with them wherever they operate. A United States that refuses to engage does not stop global AI governance from happening; it ensures that global AI governance happens without American input.
Third — and most important — some AI risks are genuinely borderless. A biological-design model that slips its safeguards can be copied across the planet in seconds. A cyber-capable agent does not stop at a customs checkpoint. These are the risks that motivated the Security Council session in the first place, and they are precisely the risks that no single country can manage alone, because the attack surface is the whole internet. One does not have to believe in a Hollywood doomsday to believe in mandatory incident reporting, shared evaluation standards, and agreed emergency protocols — the boring, technical furniture of global risk management. The administration's rhetoric conflates all of that with "centralised control," which is a category error. Reporting a model's dangerous capabilities to an international body is not the same as letting that body license American software.
What the public actually wants
There is a further complication for both camps: the public is not on either side's team. A CNN poll conducted by SSRS and reported this month found that Americans regard the growth of AI with far more fear and concern than hope and excitement — and that there is broad, bipartisan agreement that the federal government should do more to regulate it. This is a striking finding in an era when bipartisan agreement on anything is rare. It suggests that the "move fast" consensus exists mainly among investors and engineers, not among the people who will live with the consequences. And it makes the current congressional posture look stranger still: reporting this month revealed that Senate lawmakers and their staff are barred by the chamber's security rules from using anything beyond basic AI chat tools, even as they write the rules for the technology's most advanced forms. As one policy advocate put it, lawmakers are "writing rules for technology they, in many cases, never even used." That is not a recipe for wise legislation — but it is a symptom of how far the governance conversation has drifted from the public it claims to serve.
The political pressure is also building from the states. Reports this month described 23 states pushing for their own AI regulation, while the Pentagon has moved to label Anthropic a national security risk. This fragmentation is the predictable consequence of federal inaction: when Washington declines to set the rules, everyone else does, and the result is a patchwork that companies must navigate state by state and country by country. Those who oppose federal regulation on the grounds that it would stifle innovation should reckon with the possibility that the alternative — dozens of overlapping regimes — stifles it more.
A middle path worth taking
So what would a serious, non-fantastical AI governance agenda look like in 2026? It would start by separating the two problems that are constantly being fused: the geopolitical race and the safety commons. The race — who builds the most capable systems, who sets commercial standards — is legitimately a competition, and no treaty is going to end it. The safety commons — the shared interest in preventing accidents, misuse of dangerous capabilities, and cascading failures — is a coordination problem, and coordination does not require trust; it requires verification mechanisms proportionate to the risk.
Concretely, that means a few unglamorous things. Mandatory incident reporting for AI systems that cause or nearly cause serious harm, with data shared internationally the way aviation accidents are. Common, independently administered evaluations for the most capable models — the "third party evaluators" Amodei himself proposes — with results that are public by default. Minimum security standards for the weights of frontier models, because a stolen model is an ungovernable model. And a genuine research program, funded at scale, on the science of controlling advanced systems — the very work the executives say they want more time for. None of this requires a global AI government. All of it requires the United States to show up at the table it currently refuses to sit at.
What it means for the rest of the world
For audiences outside the United States, the stakes are straightforward. Most countries will never train a frontier model, but all of them will deploy AI systems, absorb AI-generated disinformation, and live with AI-enabled cyber threats. They have a direct interest in whether the most capable systems in the world are tested before release and whether their failures are reported honestly. If Washington continues to treat international AI governance as a zero-sum game, the likely outcome is not an unregulated American AI golden age but a fragmented one: American models governed by American politics, Chinese models governed by Chinese politics, and everyone else governed by whichever standards arrive first in their app stores. That is a world in which the risks the Security Council was convened to discuss are more likely, not less.
The Council's first session on AI was historic. What would make it matter is a second session — one where the proposals on the table are tested against the realpolitik the critics demand and the urgency the executives profess. The executives asked for time and coordination. The administration said no. Somewhere between those two answers lies the actual work of governance, and it will not do itself.
Sources
