The obligation covers both visible and latent signals
The law targets large, publicly accessible generative-AI systems and focuses on synthetic image, video, and audio. Providers must make a public detection tool available, offer a clear manifest disclosure option, and include latent information such as the provider, model version, creation time, and a unique identifier when technically feasible and reasonable.
The statute also restricts retention of personal information submitted to detection tools and requires licensed systems to preserve disclosure capability. A provider that learns a licensee disabled that capability must revoke the license within 96 hours.
A watermark is useful only if the ecosystem preserves it
California’s 2025 amendments add duties for large online platforms and model-hosting platforms beginning in 2027, including rules against knowingly stripping standards-compliant provenance data. Capture-device requirements follow later. That staged structure recognizes that authenticity is a chain, not a single label applied at generation.
Enforcement will reveal whether the regime creates trustworthy verification or merely a compliance badge. Providers should disclose false-positive and false-negative performance, test common edits and re-encodings, and explain when provenance cannot survive. Consumers need reliable evidence, not a new icon whose limitations remain invisible.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
California Legislature — SB 942, California AI Transparency Act California Legislature — Current Business and Professions Code, Chapter 25 California Legislature — AB 853 operative-date and platform amendments
