California Transparency in Frontier AI Act

Statute United States (California)

SB 53 is narrow by design and important anyway. Signed 25 September 2025, operative from 1 January 2026, it applies to frontier developers — those training foundation models above 10^26 integer or floating-point operations, counting subsequent fine-tuning and material modification — with the heaviest duties reserved for “large frontier developers” above a revenue threshold. That is a population you can count on your fingers.

• Publish a frontier AI framework - How the developer identifies, assesses and mitigates catastrophic risk, how it incorporates national and international standards, and how it governs the decision to deploy.
• Publish a transparency report at release - Model characteristics, intended uses, restrictions, and the catastrophic-risk assessment performed.
• Report critical safety incidents - To the California Office of Emergency Services, on a defined clock.
• Whistleblower protections - For employees raising catastrophic-risk concerns, which is the provision that will produce the first real test of the statute.
• Civil penalties to $1 million per violation - Enforced by the Attorney General.

I want to separate two things that get conflated. As a regulation of catastrophic risk, SB 53 is modest — it compels disclosure of a framework, not the adequacy of one, and a developer who publishes a thin framework and follows it has complied. As a standardisation event, it is significant, because it is the first time anyone has legally required these documents to exist in public, in a comparable form, on a schedule tied to release.

That comparability is the thing I care about. Frontier labs have published model cards and system cards voluntarily for years, and the result has been a genre with no shared schema, no required fields, no release-time guarantee, and no way to diff one against the next. Once a document is legally required at release, it acquires a version history, and a version history is what makes a claim checkable. The first time a transparency report contradicts the previous one, somebody will notice, and they will only notice because both exist.

The companion statute matters more for everyone else. AB 2013, effective the same day, requires any developer of a generative AI system made available to Californians to publish a summary of the data used to train it. That is not a frontier obligation — it is the broadest AI disclosure requirement currently in force in the United States, and it catches small companies and fine-tuners that SB 53 never touches. If you are trying to work out which single US AI provision is most likely to apply to you, it is AB 2013, not this one.

For the catalog, both point the same direction: the durable, survivable AI obligations are disclosure obligations, and disclosure obligations are the ones an API provider can actually satisfy in a machine-readable way if anyone builds the schema.