OpenAI changes course and asks California to strengthen its AI law

OpenAI is unexpectedly changing its stance on artificial intelligence regulation in California. The company previously opposed SB 53, which introduced transparency requirements and whistleblower protections for large AI companies. It is now asking California to add more safeguards to the same law.

California is meant to fill the federal gap

OpenAI explains the shift by pointing to the absence of significant federal legislation. It now supports an approach it calls reverse federalism: individual states would move in compatible directions on core protections and create the basis for a future national standard.

This is therefore about more than one isolated California rule. The company is asking for state frameworks that can align now and potentially converge into a national baseline later.

Monitoring moves into model training

One proposed amendment would require monitoring frontier models during training or evaluation for potential serious incidents. Oversight would begin before deployment, while a lab is still examining a model's capabilities and risks. OpenAI wants to avoid a scenario where 95% of issues are addressed only in the finished product.

That requirement would expand the law from general transparency toward specific controls during development. This is the stage at which companies may be able to catch a problem before a model reaches users.

Cybersecurity covers the full development lifecycle

OpenAI is also calling for stronger cybersecurity protections throughout the model-development lifecycle. Its statement refers to recent incidents that, in the company's view, show why safeguards must be updated as new risks emerge.

The proposal links model-behavior safety with the security of the infrastructure used to train, evaluate, and operate the model. Stronger requirements would therefore apply across several stages of development rather than at a single final checkpoint.

Compatible state rules could move ahead of Congress

Compatibility among similar state frameworks is a condition of the approach OpenAI is advocating. The company argues that this alignment could create the core of a later national standard even without federal legislation.

If other states move in compatible directions, SB 53 will no longer function only as a local regime. It could become one of the building blocks of a shared US baseline for the safety of frontier models.

Lilith's verdict

OpenAI first opposed SB 53 and now wants more safeguards. When a giant discovers the charm of compatible rules while invoking recent incidents, the timing is almost as interesting as the law itself.