OpenAI published a proposal today urging the United States to take the lead in setting international standards for frontier AI, with the aim of coordinating rules across countries rather than letting each one write its own. The document frames the effort around two building blocks: a way to align national and international rules, and shared ways of measuring systems and reporting incidents. The company says these standards are not licenses, pre-release reviews, or approval requirements, and that national governments would still decide how to fold them into their own legal systems.
The proposal singles out recursive self-improvement, or RSI, as an area the standards should cover. It states that fully autonomous RSI is not happening today, and that the capability should not be pursued unless it can be done safely. OpenAI also points to what it calls the Hugging Face Incident, an episode it says it has disclosed, describing it as a preview of the risks ahead while noting it was not a direct result of RSI.
Read more: OpenAI Report Uncovers Global Chinese Intimidation Campaign Using ChatGPT
The document names fragmentation between jurisdictions, the difficulty of collective action, and uneven national capacity as the main obstacles, and says they apply to both open and closed models. It argues the standards should not tilt the field toward any particular company, country, or business model, and that open-weight developers and newcomers should not be put at a disadvantage. Avoiding a concentration of power is listed as one of the goals.
The document says the standards effort should consult developers of open and closed models, outside experts, and academics as it moves forward. The proposal points to existing institutions it considers democratic, including CAISI, and describes state laws and a federal AI framework as helpful pieces of the picture. It also mentions new public-private partnerships and lists AI safety institutes in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom. CAISI set up the International Network for Advanced AI Measurement, Evaluation, and Science in 2024, according to the document.
The proposal draws lessons from aviation and financial stability, and names standards bodies including ISO, the Frontier Model Forum, the Agentic AI Foundation, and the Open Secure AI Alliance. It also points to the Appia Foundation, which it says is developing practical specifications. Labs that pursue automated AI research or other advanced capabilities, the document says, must take accountability for doing so safely.
Three goals are laid out by Sam Altman and Jakub Pachocki: building an automated AI researcher that can be iterated on for alignment, delivering scientific progress and economic growth, and giving everyone access to a personal AGI. The document cites advances in mathematics, including the Navier-Stokes Millennium Problem, as an example of what AI-enabled research could produce, and lists medicine, health care, and small businesses among the potential benefits. It also asks a question the company poses for the field: "What does good look like in the mitigation of catastrophic AI risk?".
The push for shared standards follows OpenAI's September 9 announcement that Paul Christiano joined the OpenAI Foundation Board and its Safety and Security Committee, and comes after the company's earlier policy argument that the window for action was open. It also lands months after OpenAI and Oracle marked the first building completion at the $16 billion Stargate campus in Saline Township, Michigan, a site expected to deliver 1GW of capacity as part of OpenAI's roughly $600 billion compute target.
