OpenAI wants Washington to lead on global AI standards
Two rule-writing moves landed on the same day: OpenAI asked Washington to export its standards rather than legislate a licence, and the Treasury proposed a hotline to Beijing for AI incidents.
OpenAI published a Global Affairs post on Monday calling for the United States to lead an international effort to write global technical standards for frontier AI — including standards for recursive self-improvement — and stated for the first time in writing that fully autonomous self-improvement "should not be pursued unless and until it can be done safely." The company says the decision depends on preserving human control and on informed democratic choices about benefits and risks, and warns that without care, automated research could leave humans unable to oversee research processes they no longer understand. It cites the Hugging Face incident it disclosed separately as a preview of the risk class, while noting that incident was not caused by self-improvement.
The proposed machinery is deliberately light. OpenAI wants standards set through the Center for AI Standards and Innovation and national industry bodies, built on the network of AI safety institutes already standing in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom, and on CAISI's 2024 International Network for Advanced AI Measurement, Evaluation, and Science. It names the three problems the effort is meant to answer: fragmentation, where conflicting evaluations, reporting rules and incident definitions make cross-border evidence uncomparable; collective action, where countries acting alone produce outcomes none of them wants; and uneven capacity, since frontier expertise is not spread evenly. The document is explicit about its limits — these would not be licences, mandatory pre-release review or approval requirements, and each government decides whether to fold them into its own law.
That restraint is the story. A voluntary technical baseline, adopted by governments that choose to, is the weakest form of global governance anyone has proposed this month — and it arrives from the lab with the most to gain from keeping binding rules off the table. The post also asks that the standards not advantage particular companies, countries or business models, "including by making it harder for new entrants or open-weight developers to compete," which is a commitment rivals will hold it to. We covered the domestic half of this argument earlier this month — OpenAI wants national AI rules — and CAISI's evaluator joins its board — and the RSI line lands as the pressure to go further builds: Ezra Klein calls for a ban on self-improving AI, and OpenAI's answer is a standard, not a stop.
Treasury Secretary Scott Bessent said the United States has proposed a notification mechanism with China for AI incidents that rise "up to a national security level," after roughly eight hours of talks with Vice Premier He Lifeng's team at JPMorgan's New York headquarters on Sunday. Bessent called the engagement "very successful on trade and AI" and said both sides would meet again; Reuters reports the next round is expected in about two months in Shenzhen, the city hosting November's APEC summit. China's delegation did not address reporters, so the channel exists as a US proposal Beijing has not publicly endorsed. OpenAI's post flagged the same opening, calling US–China dialogue "a positive step" at an opportune moment — worth noting, since a single lab's standards document now reads partly as a brief for the government talks happening two days later.
The trade track is moving on a separate clock. Trade Representative Jamieson Greer said Washington could support a three-to-six-month extension, with teams still working out which "non-sensitive" goods could be carved out for mutual tariff cuts. No extension has been confirmed. The honest read on the AI half is that a notification channel is the cheapest deliverable available to two governments that cannot agree on what counts as a serious incident: no thresholds published, no named contact points, no verification attached. A phone line for a category neither side has defined is still better than learning about a rogue agent from a research blog months later — but it is a channel, not a rule.
What to watch: whether the Shenzhen round two months out attaches numbers to the channel — severity thresholds, reporting timelines, who calls whom — and whether the safety-institute network actually adopts OpenAI's baseline or lets it sit as a blog post.
If two governments can't agree on what counts as an AI incident, what should the first reported one be? Tell us in the comments.
Sources: OpenAI — Building standards for the next phase of AI · Unite.AI · Reuters · Bloomberg · AP · France 24
