OpenAI is calling for international technical standards governing recursive self-improvement, a possible future process in which AI systems automate more of the research needed to build their successors. The company says fully autonomous recursive improvement is not happening today and should not be pursued until safeguards are mature.
Its proposal includes common measurement methods, incident-reporting practices and rules that preserve human oversight of automated research. OpenAI wants the United States to lead work through national safety institutes and standards organizations such as ISO. A recent United Nations scientific panel has raised related concerns about autonomous agent groups and suggested cooperation modeled on aviation, nuclear energy and cybersecurity.
The risk is not simply that one model changes itself. Automation could compress the cycle of designing, training and evaluating later systems until people have difficulty understanding or interrupting it. OpenAI warns that weak controls could produce systems that are less aligned and more dangerous.
These are recommendations from a company developing frontier models, not an agreed treaty or evidence that such a capability is imminent. Policymakers will still need definitions, independent tests and enforcement mechanisms that distinguish speculative scenarios from measurable present-day risks.