California's AI Transparency Act became operative on August 2, the same day the European Union's transparency phase began. From that date, a generative-AI provider with more than 1,000,000 monthly users or visitors in California has to publish a free tool that tells anyone whether a given image, audio or video came out of its system, offer users an optional visible disclosure, and embed machine-readable provenance data in what it generates.
What actually switched on
Only the provider tier. The statute is SB 942, rewritten by AB 853, approved by the governor on October 13, 2025. Its operative dates are staggered across four categories of company, and August 2 belongs to exactly one of them: the model builders themselves.
What did not
Large online platforms — the tier defined at 2,000,000 unique monthly users, which is where the consumer-facing detection duty actually lives — are not covered until January 1, 2027. Generative-AI hosting platforms carry the same 2027 date. Manufacturers of capture devices, meaning the cameras and phones that would stamp provenance at the moment of recording, have until January 1, 2028, and then only for devices first produced for sale on or after that day.
The date moved backwards, not forwards
The framing to watch is "California now requires platforms to detect AI content." Three of the four categories are not live, and the platform duty is 17 months away. AB 853 also accelerated nothing: SB 942 as originally passed was due to start on January 1, 2026. The amendment moved the provider tier seven months out and parked everything else in 2027 and 2028.
Same day as Brussels, different machinery
The collision with the EU AI Act is real but shallow. Brussels is phasing obligations under a single regulation with deferrals of its own; Sacramento is starting a state statute whose heaviest duties fall on companies the EU text reaches differently. What the two share is direction: provenance metadata as the compliance object, rather than detection after the fact.