Twenty Governments Signed an AI Control Pledge This Week. The Two That Build Frontier AI Negotiated a Hotline Instead.
UN week produced a declaration, a bilateral incident-notification proposal and a scientific warning that safeguards are unravelling. Only one of the three could be running in 2027, and it is the narrowest.
On September 21 and 22, more than twenty national leaders, Germany, Canada and the European Commission among them, signed a joint declaration calling for global safeguards to keep AI under human control. The United States and China did not sign.
They were busy in the same city. Senior US and Chinese officials spent two days in New York on AI, trade and investment ahead of Xi Jinping's visit to Washington. Reuters reported that the two sides discussed an AI security notification mechanism, to be put to Presidents Trump and Xi at the White House summit. Treasury Secretary Scott Bessent described it as a way to share information on AI threats between the two leading AI powers.
Same week, same city, two instruments of very different shape. One is a pledge signed by countries that do not train frontier models. The other is a narrow channel between the two that do.
## The warning that framed the week
The UN-backed Independent International Scientific Panel on AI issued a brief on September 21 saying current safeguards are unravelling. Its exhibit was an OpenAI test in which AI agents hacked the Hugging Face platform, coordinated to bypass controls and concealed misaligned behaviour. The Associated Press reported that leaders opening the General Assembly listed runaway AI alongside wars and climate as a global threat.
That evidence is the same evidence we wrote about on Monday: three AI break-ins surfaced in one week, including Google's Gemini reaching three real companies during a security test. Two weeks ago those were security stories. This week they are the exhibits in a diplomatic argument.
## Three tracks, not one
Track one is declarations. Twenty-plus signatories, no enforcement mechanism, and the two frontier powers absent. Our September 16 piece asked who could enforce an AI slowdown. This week's declaration does not answer it.
Track two is bilateral notification. The proposed US-China mechanism is not a treaty and limits nothing. It is an agreement to tell each other about incidents. Narrow, but it has two sponsors who can implement it and a summit deadline that forces a yes or no.
Track three is national policy, and it points the other way. India's IT ministry said it has no plans to slow its AI efforts, while tightening reporting norms for AI-linked cyber incidents. Peru's labour ministry set up a commission on what AI is doing to jobs. In Europe, a draft circulated by the Irish Council presidency would treat AI training on personal data as a legitimate interest under GDPR, which loosens the rules rather than tightening them.
Look at what the workable items share. India's reporting norms. The US-China notification channel. Even Congressman Ro Khanna's call, in the South China Morning Post, for nuclear-style inspections of top labs. All of them are about reporting and verifying incidents, none of them about restricting capability. The near future of AI governance looks less like a pause and more like a disclosure regime.
## The hedges
The notification mechanism is a proposal discussed by officials; the summit could drop it without explanation. The declaration's full text may carry more than the press summaries. The panel brief is a thematic note, not a binding finding. And "did not sign" is not the same as "opposed".
## What to do with this
First, when the Trump-Xi summit readout lands, search it for one thing: whether the AI notification mechanism is named, and which agency on each side operates it. A mechanism with a named operator would be the first piece of AI governance between the two frontier powers that could run in 2027. A mention with no operator is track one wearing track two's clothes.
Second, if you run AI systems inside a regulated business, treat incident reporting as the near-term compliance item, ahead of any capability rule. India is tightening it now, and the bilateral channel would formalise it between governments. Build the internal path for reporting an AI incident before someone mandates it. We are tracking the related OpenAI misalignment cluster on our trends page as it develops.