OpenAI published a piece on Monday that does two jobs at once. It states where the company stands on recursive self-improvement, and it says who should write the rules. The line that sticks is short.
«Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.»
That is not just reassurance. In the same passage OpenAI names the scenario it wants to avoid: people losing practical control of AI development because they are asked to oversee research processes they no longer understand. The Hugging Face incident, the company writes, was not caused by RSI, but it is a preview.
What OpenAI is proposing
Two building blocks.
The first is a mechanism for standards. OpenAI wants to use the existing network of AI safety institutes – Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the UK get named – and run the threads through the US Center for AI Standards and Innovation. Explicitly not licences, not mandatory prerelease review, not approval. National governments decide for themselves what they write into law.
The second block is the more interesting one. OpenAI wants shared measurements, and three of them are spelled out: how much RSI-relevant progress is happening inside a lab and how much of its research runs autonomously; which automated research steps must trigger immediate human review; and one classification for incidents, with severity levels and reporting thresholds.
The timing is not an accident
Sam Altman speaks to the UN Security Council in New York on Wednesday, and this text is what he brings with him. Over the weekend, Treasury Secretary Scott Bessent and Vice Premier He Lifeng discussed a notification channel between Washington and Beijing for AI incidents that cross a national security threshold. OpenAI’s own text calls US-China dialogue a positive step and says the upcoming talks come at an opportune time.
Who counts how much of the research the machine does
Measuring the share of autonomous research inside a company is where this gets uncomfortable. Only one lab has published that number so far: Anthropic, at 26 percent. A standard would turn a voluntary disclosure into a comparable one, which turns a marketing figure into a metric you can hold against a company’s own safety promises.
Whether governments go along is another matter. What stands out is how close OpenAI’s proposal sits to the thing Dario Amodei was attacked for two weeks ago. OpenAI just doesn’t call it a speed limit. It calls it a standard.