What Is California's AI Transparency Act — and Who Must Comply?

California’s AI Transparency Act is the first US state law that forces makers of large, publicly available generative AI tools to mark the images, video, and audio they produce as AI-made — and to give the public a free way to check. Known formally as Senate Bill 942, its core requirements became mandatory on August 2, 2026, after a legislative delay pushed the original January 2026 start date back by seven months.

What the law requires

The Act applies to “covered providers”: companies that build a generative AI system that produces images, video, or audio and that draws more than one million monthly visitors or users in California. Text-only tools — a chatbot that just outputs writing — fall outside the disclosure rules entirely.

Covered providers must build in two layers of disclosure. The first, called manifest disclosure, is a visible marker: a label or on-screen watermark that users can choose to attach to content, which once added must be “permanent or extraordinarily difficult to remove.” The second, latent disclosure, is not optional — every image, video, or audio file the system generates must carry embedded, machine-readable metadata recording which system made it, its version, and when. This kind of hidden digital watermarking follows the same logic as C2PA content credentials, an industry provenance standard several major AI and camera makers already support.

Providers must also publish a free tool, open to anyone without a login or fee, that checks whether a piece of content carries their system’s watermark. If another company licenses a covered provider’s model and strips out its disclosure features, the original provider has 96 hours to cut off that license or risk liability itself.

Who else the law will reach

The Act’s scope grows in stages. From January 1, 2027, “large online platforms” — social networks, file-sharing sites, and search engines with more than two million monthly California users — must detect and surface that provenance data to users, and platforms that host open-weight AI models may not distribute ones that skip the disclosure requirements. From January 1, 2028, manufacturers of cameras, phones, and other recording devices must let users optionally attach similar authenticity metadata to real photos and video, so genuine footage can be told apart from AI-generated content too.

Why it matters

The Act targets a narrow but consequential problem: as AI-made images, voice clones, and video become harder to tell from real recordings, neither ordinary users nor platforms have had a reliable, standardized way to check. California’s answer is to make disclosure a legal obligation instead of a voluntary feature — with real enforcement behind it. The state attorney general can seek civil penalties of $5,000 per violation, counted per day, which adds up quickly for a company operating at scale.

The law is also a test case other governments are watching. California deliberately timed its August 2026 start date to align with the EU AI Act’s rollout for high-risk systems, and several other US states have floated similar content-provenance bills. How cleanly AI companies comply, and how well the watermarks hold up against people trying to strip them out, will shape whether other jurisdictions copy California’s specific approach or write their own rules from scratch.

In the news

The law’s requirements took effect on August 2, 2026, following our coverage of that milestone.