OpenAI says fully autonomous AI self-improvement shouldn't be pursued until it can be done safely
The same post wants the US government to run the international side, a week after Sam Altman said OpenAI wouldn't wait for it.
OpenAI published Building standards for the next phase of AI on September 21, 2026, under its Global Affairs heading. It asks the United States to lead an effort with other countries to write global technical standards for frontier AI, and it wants those standards to cover AI that builds AI.
Fully autonomous recursive self-improvement isn't happening today, the post says, and shouldn't be pursued until it can be done safely. That surprised me. The first of the three goals at the top of the same page is building an automated AI researcher, and the two sit about a screen apart (OpenAI shortens the mouthful to RSI, and so does everything below).
- What
- Global technical standards for frontier AI, including for RSI
- Who should lead
- The United States
- Through
- The emerging network of AI safety institutes, the CAISI and national industry bodies
- Not
- Licenses, mandatory prerelease review, or approval requirements for AI models
What OpenAI wants written down
The mechanism runs through AI safety institutes that governments have already built, with the US Center for AI Standards and Innovation and national industry bodies doing the standard setting. Ten countries get named as having one: Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom. China isn't among them.
The standards themselves would measure how much of a lab's AI research is being done autonomously, and would say when an automated research process has to stop for a human to look at it. A third would set common severity levels and reporting thresholds for alignment incidents. OpenAI offers its own misalignment reporting framework as an early contribution to that one (it came with six reports of misaligned behavior on September 16, 2026, and we went through them).
Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.
Where the phrase came from
Pacing the frontier is Dario Amodei's phrase. It comes from an essay on his own site that asks the labs to slow down and to give outside evaluators desks and badges inside their companies, and we wrote that one up when it went out. OpenAI keeps the verb and hands it a definition of its own: keeping alignment research, and the deployment of that research, ahead of capabilities.
an automated AI researcher can also be an automated AI safety researcher
The international half is the piece Sam Altman had already marked as his government's job, a week before this. He posted it on September 14, 2026.
The world deserves confidence that American companies developing increasingly capable AI will act responsibly, especially as the trajectory of progress has steepened. Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot. We welcome a federal framework that sets consistent safety requirements for frontier AI. But we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence. Consistent rules to manage frontier risk so that we can maximize the benefits are a good idea (and we are excited by ideas like independent auditors). Years ago, companies like ours developed things like Responsible Scaling Policies and Preparedness Frameworks. Those were good for that moment, and focused primarily on the deployment of completed models, not what happens during their development process. Today's shift to focusing on safe development and evaluation will need new tools. For example, at OpenAI we now formulate explicit safety cases in advance of frontier reinforcement learning runs we expect to significantly increase capability, in addition to the safety work we have long done in advance of model releases. We hope that other companies will learn from our approaches and propose their own; we think shared standards for misalignment, monitoring, and safety will lead to better outcomes. We look forward to collaborating with our colleagues across the industry to formulate the best version of these. When we talk about “pacing”, we do not mean “stopping”. Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs. Pacing will be well worth this cost; no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring. Where we will need the help of our government is for international coordination. But first we should do what we can ourselves.
What OpenAI leaves open
OpenAI doesn't give a date for any of it. It doesn't say what should happen to a lab that skips the standards once they exist either, and the closest it comes is a line about how any lab doing automated AI research "must take accountability for doing so safely, in accordance with basic principles of self-responsibility and existing laws."
The only thing here that sits on a calendar belongs to somebody else. OpenAI says dialogue between the United States and China on these questions would be a positive step, and that upcoming talks are happening at an opportune time.