California SB 942, Explained: The AI Transparency Act

California SB 942, the California AI Transparency Act, makes large generative AI providers label their output, embed a hidden watermark in it, and offer a free public tool that can check the mark. It took effect January 1, 2026.

The law targets the companies that build the AI systems, not the businesses that use them day to day. It only reaches a 'covered provider,' a term with a specific size test we cover below.

This guide explains what SB 942 says in plain English, who counts as a covered provider, what the watermark and detection duties actually require, and how the law compares to the EU AI Act's Article 50. It is general information, not legal advice.

What Is California SB 942 (The AI Transparency Act)?

California SB 942 is a state law that requires large generative AI systems to disclose when content came from AI. It does this through two separate mechanisms plus a public tool.

The first mechanism is a manifest disclosure, a visible label a user can choose to add to AI-generated content. The second is a latent disclosure, a hidden watermark embedded in the file or text itself.

On top of both, covered providers must run a free detection tool the public can use to check whether content came from their system. The law treats the watermark and the checker as a pair. One creates the signal, the other reads it back.

• Manifest disclosure: an optional visible AI label the user can turn on.
• Latent disclosure: a hidden, machine-readable watermark embedded automatically.
• Detection tool: a free, public checker the provider must operate.
• Effective date: January 1, 2026.

Who Does SB 942 Apply To?

SB 942 applies only to a 'covered provider,' a company that makes a publicly accessible generative AI system available to more than one million people or devices in California in a 12-month period.

That threshold is high. It reaches the handful of frontier labs behind the AI tools most businesses use, such as OpenAI, Google, and Anthropic, not the businesses that build on top of those tools.

Across the dozens of sites in our own content-automation portfolio, none come close to a million monthly AI-system users, which is the ordinary case. If your business writes with AI, publishes with AI, or resells an AI feature, you are almost certainly a customer of a covered provider, not one yourself.

The practical question to ask is simple. Did your company build and ship the generative AI model, or are you a user of someone else's model? SB 942's direct duties fall on the first group.

• Covered provider test: a publicly accessible GenAI system used by 1,000,000+ people or devices in California over 12 months.
• In scope: the large labs that build and ship frontier models at that scale.
• Usually out of scope: businesses that use, resell, or build on top of someone else's AI model.
• Not a threshold that grows with your business: it is about the AI system's own reach, not your company's size.

Manifest Disclosures: The Visible AI Label

A manifest disclosure is a visible label that marks content as AI-generated or AI-altered. SB 942 requires covered providers to give users the option to add one.

The law sets a plain-language bar for the label. It must be clear, sit where a reasonable person would notice it, fit the type of content, and be hard to strip out or hide.

This is the 'seen it, know it' half of the law. A caption on an image, a tag on a video, or a note on a page can all serve as a manifest disclosure, as long as it meets that bar.

• Must be clear and conspicuous, not buried in fine print.
• Must fit the medium, an image label reads differently than a text tag.
• Must be understandable to an ordinary person, not just a specialist.
• Must be difficult to remove without notice.

Latent Disclosures: The Hidden AI Watermark

A latent disclosure is the hidden watermark half of SB 942, a machine-readable signal embedded directly in the AI-generated content. Unlike a manifest disclosure, a person cannot see it.

The law spells out what the embedded record must carry: the provider's name, the AI system's name and version, the time the content was created, and a unique identifier for that piece of content.

Covered providers must build this watermark so their own detection tool can read it back reliably, and they are expected to follow current industry watermarking standards rather than inventing a private format.

This is the same category of technique Claude uses for its own text watermark and file provenance, which we cover in our AI watermarking guide. SB 942 does not require any one method. It requires a durable, readable signal that survives ordinary handling of the content.

• What it carries: provider name, system name and version, creation timestamp, unique content ID.
• Who reads it: the provider's own detection tool, built to recognize its own mark.
• How it is built: following current industry watermarking standards, not a private one-off format.
• What it is not: visible to the reader, unlike a manifest disclosure.

The Free AI Detection Tool Providers Must Offer

SB 942 requires every covered provider to operate a free, publicly accessible tool that checks whether a piece of content carries its watermark. Anyone can use it, not just paying customers.

The tool has to accept both an uploaded file and a URL, and the law also expects API access so other software can query it programmatically.

When it finds a match, the tool returns the system's own provenance data rather than a plain yes or no, while protecting any personal information swept up in that data.

This closes the loop that the EU AI Act's Article 50 leaves more open. Article 50 requires marking; SB 942 requires marking plus a specific public tool to check the mark.

• Free and open to the public, not gated behind an account.
• Accepts file uploads, URLs, and programmatic API calls.
• Returns the system's provenance data, not a bare true/false.
• Must protect personal information contained in that data.

SB 942 vs the EU AI Act's Article 50

California SB 942 and the EU AI Act's Article 50 both push AI providers toward marking synthetic content, but they are built differently, and a company can be in scope for one and not the other.

Article 50 is broader on paper. It reaches any provider or deployer whose AI output lands in front of EU users, including small companies, and it leans on flexible techniques like watermarks, metadata, or logging.

SB 942 is narrower on who it covers but more specific on what it demands. Its one-million-user threshold excludes most companies, but the providers it does cover must ship a defined watermark payload and a working public detector, not just some detectable signal.

A large US lab selling into both California and the EU has to satisfy both laws at once, which is one reason a single global watermark, like the one Claude now ships, tends to be the practical answer rather than a region-specific one.

Enforcement, Penalties, and Effective Date

SB 942 became effective January 1, 2026. Covered providers have been expected to meet its disclosure, watermark, and detection-tool duties since that date.

The law is enforced civilly, with penalties reported at $5,000 per violation, brought by the California Attorney General or a local prosecutor. Verify the current figure against the bill text before relying on it, since enforcement guidance can be updated.

SB 942 also reaches third-party licensing. If a covered provider licenses its model to another company, the contract must require that licensee to keep the disclosure features working, and the provider must cut off the license within 96 hours of learning a licensee disabled them.

That licensing clause matters more than it looks. It stops a covered provider from meeting its own duties directly while quietly licensing a stripped-down, unmarked version of the same model to a partner.

• Effective date: January 1, 2026.
• Penalty: reported at $5,000 per violation (verify current figure).
• Who enforces it: the California Attorney General or a local prosecutor.
• Licensing rule: licensees must keep disclosure features on, or the license gets pulled within 96 hours.

What Businesses Should Do Now

Most businesses reading this are customers of a covered provider, not one. The real work is knowing which AI systems you rely on and whether those vendors are keeping up their end.

Start by listing which generative AI tools touch your public-facing content, then check that each one large enough to be a covered provider still exposes its manifest disclosure option and its detection tool.

If you publish AI-generated content to a California audience, turning on the visible AI label where your vendor offers one is good practice even before any law forces your hand, since it builds trust with readers.

For the deeper mechanics of how a watermark like this actually works once it is embedded, see our AI watermark detector guide and the wider AI watermarking guide. When you want your AI stack mapped against rules like SB 942 and Article 50 together, that is the kind of inventory we build in an AI workflow audit.

• List every generative AI tool that touches content your California audience sees.
• Confirm large vendors still expose a manifest-disclosure option and a working detector.
• Turn on visible AI labels where offered, even where not strictly required.
• Revisit the list whenever you add a new AI tool to a public workflow.