The California AI Transparency Act

The California AI Transparency Act (SB 942) follows the EU AI Act Article 50(2) in mandating AI transparency labels. Unlike the EU AI Act, it covers a much broader set of companies:

In September, the minimum-user requirement for generative AI providers was removed and a way to distinguish between AI-generated and AI-modified was added. The fines are $5000 per “violation” where each day counts as a separate violation.

The main requirements are:

• (1) to be “compliant or interoperable with widely recognized industry standards” a.k.a. C2PA metadata
• (2) to add a latent disclosure that is “permanent or extraordinarily difficult to remove” a.k.a. watermarks
• (3) a free verification tool (which can be rate-limited for security or misuse concerns)

In comparison to the EU AI Act Article 50(2):

• licensed usage of models is not explicitly covered (whereas the EU AI Act includes “under its own name or trademark”
• internal usage is not explicitly covered (“publicly accessible” in California vs. “putting into service” or “own use” in the EU)
• only covers image, audio, video (no text)

Notably, per § 22757.3(a)(1), this latent disclosure must lead to information that includes:

• the name of the AI provider
• the name and version of the AI model
• the time and date of creation
• whether it is AI-generated or AI-modified
• a unique identifier

At the same time, the verification tool cannot disclose private information, unless consent has been given “clearly and conspicuously in plain language.”

Reading this literally, it appears that a per-content watermark (such as Trufo’s PawPrint in provenance mode) is required, and a per-model watermark (such as Google’s SynthID¹ or Trufo’s PawPrint in “minimal for EU” compliance mode) is not sufficient. Furthermore, any inclusion of CAWG identity assertions (or similar disclosures of identifying information) in the C2PA manifest cannot be done haphazardly.

¹Note that SynthID may carry unique identifiers that lead to the requisite information, but we were unable to determine if this was the case from the public verification endpoint.

In January, requirements for platforms come into effect.

For online platforms (e.g. social media, search engine, file-sharing, mass-messaging) with more than 2 million users, they must:

• (1) detect provenance data
• (2) show (i) if provenance data is present, (ii) AI-generated vs. AI-modified vs. camera-captured, (iii) the name of the AI system or capture device
• (3) avoid stripping provenance data

The requirements for AI hosting platforms covers all downloads of generative AI systems. Read literally, any platform that allows users in California to download system code or model weights must ensure that the downloaded material meets all the requirements for generative AI providers. It is unclear what this means for the downloading of raw model weights, as the model weights alone are not sufficient to embed the requisite latent disclosures.

In addition to regulating AI-generated or AI-modified content, California further regulates “real” captured content.

The main requirement is to be “compliant with widely adopted specifications adopted by an established standards-setting body” a.k.a. C2PA metadata, so watermarks are not required. This ensures that the provenance of captured content can be proven and widely understood. For example, this would be useful in identifying real-life news or presenting evidence in court.

Similar regulations are being phased into action around the world:

• EU AI Act Article 50(2): from August 2, 2026 (December 2 for existing systems), providers must mark AI outputs, including text, in a machine-readable format; fines up to €15,000,000 or 3% of worldwide annual turnover.
• EU AI Act Article 50(4): deployers must disclose deepfakes, and AI-generated text published on matters of public interest.
• China: the Labeling Measures and standard GB 45438-2025, in effect since September 1, 2025, require visible labels plus a metadata label, which platforms must check.
• South Korea: the AI Basic Act, in effect since January 22, 2026, requires generative AI outputs to be marked, visibly or machine-readably; correction orders come first, then fines up to KRW 30M.
• India: the IT Rules amendment, in effect since February 20, 2026, requires visible labels plus permanent metadata on synthetic audio and visual content; non-compliant platforms lose safe-harbor protection.

If you are ready to build, head over to the Trufo Dashboard to get started. If you are still figuring out what is best for your use case, schedule a free consultation with Trufo to chat with an expert in the space.

Official California sources for the Act:

• California AI Transparency Act (BPC §§ 22757–⁠22757.6): the current code text.
• SB 942 (2024): the original Act.
• AB 853 (2025): the delay to August 2, 2026, and the platform and device duties.
• SB 1000 (2026): the urgency statute.
• AB 2713 (2026): the large online platform text from January 1, 2027.