California's Disclosure Rules
California has enacted state AI requirements that force large model developers to publish safety frameworks, report critical incidents to authorities and disclose data sources used in training. The rules, effective from January 1, increase emphasis on data‑source disclosure as a governance lever for model builders and their vendors. Those statutory disclosure expectations create a stronger market incentive for provenance and auditable labeling practices. (english.news.cn)
California now requires two kinds of artificial intelligence disclosures: big model developers must publish safety frameworks, and generative artificial intelligence developers must describe the data used to train their systems. (gov.ca.gov) The safety law is Senate Bill 53, signed by Governor Gavin Newsom on September 29, 2025, and it took effect on January 1, 2026. It applies to “large frontier developers” and requires them to post a frontier artificial intelligence framework on their websites. (gov.ca.gov) That framework must describe how a developer uses standards and best practices, tests models, manages risks, handles security, and responds to “critical safety incidents.” Before or when deploying a new or substantially modified frontier model, the developer must also publish a transparency report. (legiscan.com) If a critical safety incident creates an imminent risk of death or serious physical injury, the law requires disclosure to an appropriate authority within 24 hours. Large frontier developers also must review and update their framework at least once a year and publish material changes within 30 days. (jdsupra.com) The data-disclosure law is Assembly Bill 2013, signed on September 28, 2024, and also effective January 1, 2026. It requires developers of generative artificial intelligence systems or services released on or after January 1, 2022 to post training-data documentation before making them available to Californians. (goodwinlaw.com) That documentation must give high-level summaries of datasets and say where the data came from, including whether it was bought, licensed, publicly available, or obtained in some other way. The law also calls for disclosures about whether personal information is included and whether copyrighted material or data under an open license was used. (regulations.ai) California moved in this direction after Newsom vetoed Senate Bill 1047 on September 29, 2024, rejecting a broader model-safety bill that many technology companies opposed. His administration later backed a narrower disclosure-based approach after publishing a state frontier artificial intelligence policy report in 2025. (ciodive.com) (gov.ca.gov) Supporters of Assembly Bill 2013 told the Senate Judiciary Committee in 2024 that public documentation would give consumers, workers, and rightsholders a clearer record of what went into a model. Business groups opposed the bill there, arguing the requirements were broad and could expose sensitive information. (sjud.senate.ca.gov) The result is a California rulebook that leans on publication and reporting instead of direct preapproval. Since January 1, 2026, developers selling into California have had to show more of their safety process and more of their data trail in public. (fpf.org) (goodwinlaw.com)
