California’s AI Transparency Act (SB 942): the Provenance Requirements

SB 942 is the first US statute to require a generative AI provider to ship both a marking mechanism and a tool for checking it. It applies to very few companies directly. It reaches a great many indirectly, through a licence-revocation rule that most summaries do not mention.

Who is a covered provider

The California AI Transparency Act, enacted by Senate Bill 942 and signed on 19 September 2024, adds a chapter to the Business and Professions Code beginning at section 22757. Its obligations attach to a covered provider: a person that creates, codes or otherwise produces a generative AI system that has over one million monthly visitors or users and is publicly accessible within California. The bill record is at California Legislative Information.

Two things about that threshold matter more than the number. It counts users of the system, not customers of the company, so a model exposed through a popular consumer surface can cross it while the company’s enterprise business is small. And it is not a revenue or headcount test, so a well-used free product is covered where a large but narrow enterprise vendor is not. The threshold is also a moving line rather than a status: crossing it is a compliance event.

The three duties

A free detection tool. A covered provider must make available a tool, at no cost, that lets a user assess whether image, video or audio content, or content containing those, was created or altered by the provider’s generative AI system. The tool must output any system provenance data detected in the content, must not output any personal provenance data, must support a user uploading content or providing a URL, and must be accessible programmatically through an application programming interface. The statute also requires the provider to collect feedback on its efficacy and to make an affirmative representation about it.

The prohibition on outputting personal provenance data is the part worth pausing on. The tool is meant to answer “was this machine-generated”, not “who made this”. A design that returns the account that produced an image would fail the Act rather than satisfy it.

Latent disclosure. Content the system creates or alters must include a latent disclosure — provenance information embedded in the content itself, detectable by the provider’s own tool. The required elements are, in substance, that the content is AI-generated, the name of the provider, the name and version number of the system that produced it, and the time and date of creation or alteration, together with a unique identifier. The obligation is qualified: it applies to the extent technically feasible and reasonable, and the disclosure must be permanent or extraordinarily difficult to remove, to the extent it is technically feasible.

Those qualifiers are doing a lot and they are honest ones. No embedding technique currently survives arbitrary re-encoding, cropping and screenshotting, and the drafters plainly knew it. The consequence is that the standard is comparative — what is technically feasible moves as the field moves — rather than absolute. How the embedding is actually done, and how the C2PA content credentials specification maps onto it, is covered in the provenance mechanism page and in the C2PA compliance page.

Optional manifest disclosure. The provider must offer the user the option of a manifest disclosure — a visible label on the content saying it is AI-generated — which must be clear, conspicuous, appropriate for the medium, and understandable to a reasonable person. This one is at the user’s election. SB 942 does not require every generated image to carry a visible watermark; it requires the provider to make that available.

The 96-hour licence revocation rule

This is the provision that gives the Act reach beyond the handful of companies over the threshold. Where a covered provider licenses its generative AI system to a third party, and the licensee makes a modification such that the system no longer includes the required disclosures, the provider must revoke the licence — and the statute sets a ninety-six hour window running from when the provider knew or should reasonably have known of the modification. The licensee, on revocation, must cease using the system.

Read from the licensee’s side, that is a hard termination risk attached to a technical change. Stripping metadata to reduce file size, re-encoding through a pipeline that discards embedded provenance, fine-tuning in a way that disables a marking step — none of these need to be deliberate evasion to trigger the analysis. Anyone building on a licensed model from a covered provider should know where in their pipeline provenance data can be lost, and should expect the licence agreement to say something about it.

Dates, and the amendment that moved them

As enacted, SB 942 was to operate from 1 January 2026. It was amended before that date by AB 853 in the 2025 session, which postponed the covered-provider obligations into 2026 and extended the Act’s scheme to other actors — large online platforms, which are directed at retaining and surfacing provenance metadata in content they distribute, and capture devices, on later dates. The bill record is at California Legislative Information.

Enforcement and what it does not cover

The Act provides for a civil penalty of five thousand dollars per violation, with each day of a continuing violation counted as a separate violation, recoverable in an action brought by the Attorney General, a city attorney or a county counsel. There is no private right of action. Because the penalty accrues daily, the exposure from a systemic failure to embed disclosures is not bounded in any practical sense, which is a different risk profile from a one-off fine.

Three limits are worth stating explicitly, because the Act is often described as a general AI labelling law and it is not. It reaches image, video and audio content; text is not within the latent disclosure duty, which reflects the fact that no reliable text watermarking exists. It binds providers, not users: an individual who generates an image and posts it is not regulated by this chapter, though other California statutes may reach them. And it says nothing about accuracy or harm — a correctly labelled deepfake satisfies SB 942 entirely, and whether it is otherwise lawful is a question for the digital replica law or the election deepfake statutes rather than this one.

For the EU analogue, which is structured differently — a machine-readable marking duty on providers of generative systems, plus a labelling duty on deployers of deepfakes — see the AI Act marking duty. A provider serving both markets needs one mechanism that satisfies both, and the two texts do not use the same vocabulary for it.