California's AI Transparency Act, explained
California's AI Transparency Act, originally SB 942 and amended by AB 853 (signed October 13, 2025), is the first US law to put provenance duties on platforms. Covered providers of generative systems with more than a million monthly users must offer a free detection tool and include disclosures in output from August 2, 2026. From January 1, 2027, large online platforms must detect provenance data in the content they distribute, let users inspect it, and must not knowingly strip it, and generative AI hosting platforms may not knowingly offer systems that lack the required disclosures. Capture device makers follow on January 1, 2028. Individuals who edit or paraphrase text are not covered.
Provider duties from August 2, 2026
A covered provider must make available a free AI detection tool that lets a user check whether content was created or altered by the provider's system, and must include disclosures in generated content: a latent disclosure that carries provenance information, and the option of a manifest disclosure that is visible. The threshold is more than one million monthly users of a publicly accessible system in California. The date was moved from January 1, 2026 to August 2, 2026 by AB 853, aligning it with the EU's Article 50 start.
Platform duties from January 1, 2027
AB 853 added large online platforms, defined by monthly user counts, to the Act. From January 1, 2027 they must detect provenance data that complies with widely accepted standards in the content they distribute, provide a way for users to inspect that data, and must not knowingly remove it. Generative AI hosting platforms, meaning services that make generative systems available, may not knowingly make available a system that lacks the disclosures the Act requires. The platform provisions are what make this law relevant to anyone who distributes content: the duty to preserve provenance now sits with the pipe, not only with the source.
Capture devices from January 1, 2028
Manufacturers of cameras, phones and recorders sold in California must offer users the option to attach provenance data to what the device records. That is the authentic-content half of the provenance system: if genuine photographs and recordings carry signed credentials, the absence of credentials becomes informative.
How it fits with the EU
The EU puts marking duties on providers and disclosure duties on deployers, and asks signatories of its Code of Practice to keep marks non-removable. California puts detection and disclosure duties on providers, and preservation duties on platforms. Between them, the two regimes cover the whole path from generation to distribution, which is the only way provenance can work when any single link can strip it. Neither regime reaches the individual who runs a paraphrase pass, which is why the technical arms race continues alongside the law.
Reading the text
The chaptered text of AB 853 is on the California Legislature's site, linked from the policy section. Definitions of covered provider, large online platform and generative AI hosting platform are in the amended Business and Professions Code sections, and the dates above come from the bill's operative clauses. Penalties attach per violation and per day. Nothing here is legal advice; it is a map of where the duties sit and when they start.
Questions people ask
Who does the California AI Transparency Act cover?
Covered providers of generative AI systems with more than a million monthly users accessible in California, and from 2027 large online platforms, generative AI hosting platforms, and from 2028 capture device manufacturers.
Does the California law make it illegal to remove a watermark?
It bars large online platforms from knowingly stripping provenance data from January 1, 2027. It does not create a duty for individual users.
What did AB 853 change?
It delayed the provider date to August 2, 2026 and added the platform, hosting platform and capture device duties.