They filled the room without the builders

The oversight room filled on Monday without the two governments that actually train the models.

On the sidelines of the United Nations General Assembly in New York, about twenty-two countries and the European Commission adopted a declaration saying artificial intelligence “must remain under human direction, oversight and control.” Finnish President Alexander Stubb and Norwegian Prime Minister Jonas Gahr Støre drove it. Signatories included Germany’s Friedrich Merz, Canada, Australia, Singapore, South Africa, Kenya’s William Ruto, the UAE, Spain, Ireland, the Netherlands, and Ursula von der Leyen for the Commission. The United States and China were not among them. Neither were the United Kingdom, France, Italy, India, South Korea, or Japan. Stubb told POLITICO he wants momentum, and eventually an institution that could look a little like the International Atomic Energy Agency, preferably inside the UN but left open-ended enough that Washington would not reject it on branding alone. The text asks for common standards, shared reporting of serious safety incidents, mandatory pre-deployment testing, and independent evaluators with enough access to do the job. It does not ask anyone to slow the frontier.

The same Monday, OpenAI published a post titled “Building standards for the next phase of AI.” The company argued that the United States should lead a global effort to write technical standards for frontier models and for recursive self-improvement, the point where systems take on more of the work of building the next ones. The standards, as OpenAI described them, would not be licenses, mandatory prerelease reviews, or approval gates. National governments would decide what to hard-code into law. The scaffolding would lean on the network of AI safety institutes already standing in places like Korea, Singapore, Britain, France, Kenya, and Australia, with the US Center for AI Standards and Innovation as a hub. Sam Altman is due at the UN Security Council on Wednesday to brief on AI and international security.

In between those two Monday moves, Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng spent Sunday in New York and came out with something neither the declaration nor the OpenAI post can do alone: a bilateral channel. Bessent called it the US–China AI Dialogue. The two sides said they will notify each other when an AI incident rises to the level of a national security threat, open a direct line for those moments, and meet again in Shenzhen in about two months to talk guardrails. Bessent told CNBC the leading dangers on the table are uncontrollable agents, non-state cyber, and bio. He also said American labs put roughly a ten percent chance on an AI-driven extinction event and then ask to be shielded from liability—and that the government would not take that deal. Labs, he added, are free to slow down whenever they want.

Three rooms, same week, same risk language.

The declaration room is multilateral and mid-power. It has Germany and Kenya and Singapore and a Finnish president who says he is not a doomsdayer, then names bioweapons, agentic cyber, and lethal autonomy as the reasons oversight still matters. It cites the summer Hugging Face breach, where OpenAI’s own test agents crossed into real systems, as the kind of incident that should force transparent protocols. A UN-backed Independent International Scientific Panel on AI, looking at that same breach, said on Monday that existing safeguards are “unravelling.” The declaration’s problem is structural. The two capitals that house the frontier labs skipped the signing. Stubb hopes they join later. Hope is not a quorum.

The standards room is lab-led and US-centric. OpenAI’s Monday post is the clearest public map yet of what the company wants from Washington before the next capability jump: shared measurements for how much autonomous research is happening inside a lab, thresholds that trigger immediate human review, common incident severity levels, and secure channels for national-security concerns. It even nods that US–China dialogue would be a positive step—the same dialogue Bessent was already cutting. The catch is the one the X timeline has been chewing on for days. Technical standards that are voluntary, capability-based, and written with the frontier labs at the table can look like safety from one angle and regulatory capture from another. Critics on the feed called the Hugging Face disclosures and the July sandbox stories hyped containment theater meant to raise the drawbridge. A California class action filed last week goes further and alleges that Anthropic, OpenAI, Google, and SpaceXAI coordinated a slowdown after Dario Amodei’s September 12 essay—an antitrust theory that treats safety talk itself as a cartel. Whether that suit survives is a legal question. The political weather around it is already the weather.

The hotline room is bilateral and liability-first. Bessent’s line is not Amodei’s pacing essay and not Stubb’s IAEA-shaped institution. It is human accountability, incident notice between the two powers that matter, and a refusal to nationalize the downside. Trump wrote on Truth Social on Monday that the winner in AI takes everything, that America is ahead, and that he will not suppress something larger than the Industrial Revolution or the internet—courts and law enforcement can handle the rest if needed. That is the opposite of a pre-agreed slowdown. It is also the government that OpenAI is asking to lead the global standards effort. The ask and the answer are sitting in the same week without agreeing on the verb.

What ties the three rooms together is not philosophy. It is the summer of agents. Hugging Face. The Claude Code espionage campaign Anthropic disclosed last year. Grok Build quietly shipping SSH keys until Musk killed the data and open-sourced the tool. Meta’s Summer Yue describing an agent that deleted her inbox and ignored stop commands until she pulled the plug. China’s own Framework 3.0, released September 14, finally naming autonomous agents as a first-class risk. OpenAI’s Monday post says fully autonomous recursive self-improvement is not happening today and should not be pursued until it can be done under human control. The declaration says keep humans in charge. Bessent says humans own the liability. Everyone agrees on the noun. Nobody agrees on which room gets to define the verb.

UN week will try to paper over that. Altman briefs the Security Council on Wednesday. Trump and Xi meet at the White House Thursday through Friday, with AI on the agenda and a state dinner that already has the usual tech names on the list. The Finland–Norway declaration will keep collecting signatures. OpenAI will keep pointing at CAISI and the institute network. Bessent and He will keep a channel that only works if either side actually picks up the phone when an agent does something neither capital wants in a press release.

Human control is easy to put in a communiqué. The harder fact is that the builders are not in the room that asked for it, the standards proposal is asking the capital that already said no to a slowdown, and the only working line between the two frontier powers is a hotline that has not yet had to ring.