Loading page…
The company wants enforceable requirements aimed at the most capable systems, while asking states to keep building safeguards before Congress acts.
Listen to this story
OpenAI is shifting from voluntary AI safety commitments toward a policy mix of federal mandates, California enforcement infrastructure, and industry standards. It wants requirements triggered by model capabilities and risks, limited to frontier labs rather than applied broadly to smaller developers or open-model researchers. Four California bills now have OpenAI’s support and are headed to Governor Gavin Newsom, while Congress is urged to act before December.
OpenAI proposes common testing, independent frontier-model assessments, cybersecurity rules, and serious-incident reporting.
Its federal framework would target a small number of well-resourced frontier labs, not all AI developers.
California bills cover auditor accreditation, AI-risk assessment, companion-chatbot protections, and gene-synthesis screening.
OpenAI is calling for mandatory U.S. safety rules for the most powerful AI systems while backing four California measures intended to strengthen oversight now. Its central argument is that obligations should be tied to what an AI model can do, rather than the size of the company that developed it.
In its policy statement, OpenAI said voluntary commitments are no longer enough and urged Congress to establish capability-based national regulation before it adjourns in December. The proposed framework includes common testing standards, independent assessments of frontier models, stronger cybersecurity protections and reporting requirements for serious safety incidents.
OpenAI says federal requirements should cover only the handful of well-resourced labs building the most capable systems. It also argues that frontier-safety policy should not become a way to restrict open models, which it says can serve needs for local deployment, control, security or data residency.
Until Congress acts, OpenAI says states should keep building the safety infrastructure that federal lawmakers could later adopt. The company formally endorsed four California bills that have passed the legislature and are headed to Governor Gavin Newsom, saying it had reconsidered its stance on some measures after recent capability advances.
OpenAI also wants industry-led standards for monitoring frontier systems, particularly as models gain autonomy, access more powerful tools and operate for longer periods. It says those voluntary standards should complement, not replace, mandatory federal safeguards and democratic oversight.
The announcement creates no new legal requirement. It instead places OpenAI behind a policy path that combines federal rules, state action and shared lab practices. Whether lawmakers accept its capability-based boundary will determine who must meet the resulting safety bar.
Story updates
OpenAI wants the United States to lead an international effort to set safety and security standards for AI development, warning that a fragmented global system could otherwise take hold. The appeal comes as Chief Executive Sam Altman is expected to brief the U.N. Security Council this week.
OpenAI’s proposal goes beyond broad cooperation. The company says governments should pursue compatible approaches to measuring AI capabilities, managing risk, preserving meaningful human control, and deciding when development should slow or stop. It says the U.S. is positioned to lead because its AI industry is at the technical frontier.
We should not pursue it unless and until it can be done safely.
The company is not proposing that voluntary industry work replace government rules. OpenAI says it will work with other frontier labs on voluntary standards, but maintains that those efforts should complement mandatory federal safeguards and democratic oversight. Its policy blueprint also includes shared measures for tracking progress toward recursive self-improvement.
Measuring the capabilities of increasingly powerful AI systems.
Managing risks while preserving human control.
Setting shared conditions for slowing or stopping development.
OpenAI distinguishes between AI assisting parts of research and fully autonomous recursive self-improvement, where an AI system independently drives successive generations of more capable AI. It says that latter form is not happening today and should not be pursued until it can be done safely.
OpenAI linked the argument to a July breach involving unreleased models under testing and Hugging Face, which it described as a warning of risks that could become more severe without robust safeguards and alignment. The company said the incident was not a direct result of recursive self-improvement. France, which holds the Security Council’s rotating presidency, is organizing Altman’s expected briefing.
openai.comEditorial analysis
OpenAI’s position is notable less for calling for safety rules than for the line it wants lawmakers to draw. The company argues that requirements should follow a model’s capability and risk, rather than a developer’s size or the mere fact that a model is openly released. That approach could focus scrutiny on a small number of frontier labs, but it also makes the definition of frontier capability central to any future law. The immediate test is whether Congress takes up a targeted federal framework, and whether California’s assessment and auditor measures provide a workable model in the meantime.
Loading discussion...
Join the conversation
Explain which level of government should lead.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.