The rules would increase with model capabilities and apply only to the small group of companies building the most powerful systems.
OpenAI called on Congress to impose mandatory national safety requirements on companies developing the most powerful AI models.
OpenAI said the rules should cover only the handful of well-funded companies developing the most capable systems, including OpenAI. Startups, smaller developers, and researchers working with less powerful models would not be covered.
Under the proposal, major developers would have to test their most powerful models before release and submit them for independent assessments. They would also face cybersecurity requirements and rules for reporting serious incidents.
OpenAI wants developers to watch for models that break rules during development or testing, such as entering another organization’s computer systems without permission to complete an assigned task. If a model bypasses that organization’s security and accesses, changes, or destroys its systems or confidential information, the developer must notify the affected organization in writing.
The company said federal rules are becoming more urgent because current models are beginning to help researchers build future models faster. The concern is that each new model could eventually help create an even more capable successor, accelerating development faster than companies can test new systems and add safeguards.
OpenAI said this self-reinforcing process is not happening today. However, it called for government standards to track progress toward that ability and determine when development should slow or stop.
The proposal builds on an earlier OpenAI plan for federal oversight. The company said government rules and independent reviews would replace the current system in which leading AI developers largely set and assess their own safety practices.
OpenAI also endorsed California Senate Bills 813 and 1119 and Assembly Bills 1405 and 1864. Collectively, the measures regulate AI auditors, add protections for children using companion chatbots, and require gene-synthesis companies to screen orders for potentially dangerous materials.
California Gov. Gavin Newsom has since signed Senate Bill 813, Assembly Bill 1405, and Senate Bill 1119. Assembly Bill 1864 remained before the governor as of Sept. 15.
OpenAI said it expects to support federal proposals now being developed in Congress but did not name them. Congress would have to determine which models trigger the rules, who enforces them, and what happens when a developer fails to comply.
