TL;DR
OpenAI has called on the United States to lead an international effort to write technical standards for frontier AI, with a particular focus on recursive self-improvement (RSI), where AI systems increasingly build their successors. It proposes working through the network of national AI safety institutes, the UK’s among them. The company is explicit that these would be standards, not licences, mandatory pre-release review or approval regimes.
What OpenAI proposes
The post says fully autonomous RSI “is not happening today”, and that it should not be pursued until it can be done safely. Without care, OpenAI warns, people could lose practical control of research processes they no longer understand. It cites the Hugging Face incident it disclosed as an early sign of the risks, while noting that episode was not caused by RSI.
Two components are described as essential. The first is a mechanism for national and international standards to fit together, using safety institutes in countries including Australia, Canada, Japan, India and the UK, coordinated through America’s Center for AI Standards and Innovation. National governments would choose whether to write the results into law.
The second is shared measurement and incident reporting. OpenAI suggests standards for how much autonomous research is happening inside a company, for when automated research must trigger immediate human review, and for classifying and reporting incidents by severity. It points to its own misalignment reporting framework as an early contribution.
How it fits the week’s other proposals
The post landed the same day 22 countries at the UN backed a declaration on keeping AI under human control. Both call for common standards and shared incident reporting, but they diverge on who leads. The declaration floats a new international institution, which its lead organiser would place inside the UN; OpenAI wants Washington in charge, arguing the US can either set the global framework or watch a fragmented one form around it.
It also sits alongside OpenAI’s earlier request to Congress for binding safety rules. The new proposal is lighter at the international level, leaving enforcement to each country.
Looking forward
For the UK, being named in the institute network offers the AI Security Institute a technical role in setting standards. In our view, a track that needs no new legislation also fits a government that has let its planned safety law lapse, though standards with no domestic backing would bind no one here. OpenAI also welcomes US-China dialogue on these issues and notes talks are imminent.