OpenAI reverses course and backs binding US rules for frontier AI

A regulatory reversal

OpenAI has called for binding federal safety requirements for the most capable artificial-intelligence systems, a change from its earlier resistance to tighter mandatory controls. Euronews reported that the company wants national rules tied to what advanced models can do rather than broad obligations applied to every developer. The proposal includes shared testing standards, independent assessments, stronger cybersecurity requirements and compulsory reporting of serious safety incidents.

Who would be covered

The company said the proposed duties should apply to the small group developing frontier models, while shielding startups, smaller developers and researchers whose systems remain far below that capability threshold. OpenAI also backed four California bills even as it continued to argue that a single federal regime would be preferable to a patchwork of state laws. The change is therefore both substantive and strategic: it accepts mandatory controls while seeking a narrow and nationally uniform scope.

Washington currently relies on cooperation

The existing White House approach is materially different. A June executive order directed federal agencies to create classified cyber-capability benchmarks and a voluntary process through which developers can offer the government early access to covered frontier models. The order explicitly stated that it did not create mandatory licensing, preclearance or permitting for developing or releasing new models. OpenAI's intervention places mandatory obligations back into a policy framework designed around voluntary engagement.

Europe provides a regulatory comparison

The European Union has already moved into enforcement. European Commission guidance says the AI Office can demand information, evaluate general-purpose models, require access for testing and order risk-reduction measures. Certain provisions became enforceable in August, including safety and security obligations for the most advanced general-purpose models. That structure shows what legally backed oversight can look like, although OpenAI's US proposal is not a request to copy the EU law wholesale.

Why it matters

A large model developer publicly accepting binding duties changes the lobbying balance in Congress. It may make bipartisan rules easier to frame around frontier capability, but disagreements remain over thresholds, independent evaluators, confidential model access, liability and federal pre-emption of state law. The company also has a commercial interest in rules that smaller competitors may find harder to satisfy, making the precise definition of a frontier developer central to the debate.

What comes next

The immediate question is whether lawmakers convert the principles into bill text before the congressional calendar closes. Any credible proposal must specify which incidents require disclosure, who performs external assessments, how sensitive model information is protected and what penalties follow non-compliance. Until legislation passes, the federal system remains voluntary; OpenAI's announcement is an advocacy decision, not a new legal requirement.