California enacts SB 53
Newsom signed the narrower successor to the bill he had vetoed a year earlier, requiring frontier developers above set revenue and compute thresholds to publish safety frameworks and report incidents.
- Government & policy
- Safety & alignment
- Courts & copyright
- Major
Governor Gavin Newsom signed SB 53, the Transparency in Frontier Artificial Intelligence Act, a year to the day after he had vetoed its predecessor, SB 1047. Authored again by state senator Scott Wiener, the new bill was narrower in scope and framed as transparency regulation rather than a liability regime.
SB 53 applies to “large frontier developers” — companies that have trained or begun training a model using more than 10²⁶ floating-point operations and whose annual revenue, together with affiliates, exceeds $500 million. Covered developers must publish a frontier AI framework describing how they manage catastrophic risk, incorporating recognised standards and industry practice; before deploying a new frontier model, they must publish a report giving contact information, release date, supported languages, intended uses and a summary of catastrophic-risk assessment. They must report critical safety incidents to California’s Office of Emergency Services within 15 days, or 24 hours where there is imminent danger, and submit quarterly summaries of internal risk assessments. The attorney general can seek civil penalties of up to $1 million per violation.
Where SB 1047 would have mandated a shutdown capability, third-party audits and a new state regulatory body — provisions Newsom cited when he vetoed it as regulating model size rather than deployment risk — SB 53 largely codifies disclosure obligations that the largest labs, including OpenAI and Anthropic, had already adopted voluntarily through their own published safety frameworks. Supporters presented that as the point: putting existing voluntary practice into statute, with penalties, rather than inventing new technical requirements. It left open the question that dogged its predecessor — whether disclosure without a compute-threshold trigger for external audit would be sufficient if a frontier developer’s own risk assessment proved wrong — and it arrived alongside a parallel federal effort to preempt state AI legislation altogether, which SB 53’s drafters had tried to anticipate by limiting its reach to reporting rather than technical mandates.