The proposal, published September 21, asks the United States to work with other countries on technical standards for frontier models and automated AI research, including recursive self-improvement (RSI). It is a policy proposal, not an adopted international protocol or a binding regulatory regime.
OpenAI says the proposed standards could establish shared baselines for capability measurement, risk assessment and safeguard sufficiency. They could also define how to classify, track, report and respond to alignment and automated-research incidents. Human oversight would remain part of the design, including criteria for when an automated research process should trigger immediate review.
The company explicitly says the standards would not be licenses, mandatory prerelease reviews or approval requirements for AI models. National governments would decide whether and how to incorporate them into their own legal systems. OpenAI also calls for consultation with open- and closed-model developers, independent experts and academia so that the process does not favor a particular company or business model.
OpenAI recommends using existing institutions, including national AI safety institutes, national industry bodies and the U.S. Center for AI Standards and Innovation (CAISI), rather than creating a new global regulator. It also proposes secure channels through which governments and critical-infrastructure operators could share vulnerabilities, national-security concerns and best practices.
The proposal appeared as U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng discussed a possible U.S.-China AI dialogue about notifying one another when incidents reach national-security risk levels, Axios reported. That discussion is not an agreed notification protocol, and no government adoption of OpenAI’s proposal was announced.
For AI governance and engineering teams, the immediate consequence is a clearer list of interfaces that future cross-lab standards would need to address: comparable evaluations, oversight of automated research and incident disclosures that cover both pre-deployment testing and real-world use. Whether those ideas become shared practice depends on governments and other labs, not on OpenAI’s proposal alone.