California didn't wait for Washington. While the federal government debates a national approach, the state already requires large generative AI providers to digitally sign every image, video, or audio clip they generate.
The AI Transparency Act (SB 942) became operative on August 2, 2026, after a delay introduced by AB 853. Covered providers must offer users the option to add a visible AI disclosure to their generated content, and embed a hidden, machine-readable provenance watermark in AI-generated images, video, and audio. They must also make a free, publicly available AI-detection tool that meets certain accessibility criteria.
The law covers “covered providers”: operators of publicly available generative AI systems with more than one million monthly visitors or users in California — the threshold excludes small or niche tools, but covers any platform operating at real scale. The disclosure obligations (visible and hidden) apply only to images, video, and audio generated or substantially altered by AI — AI-generated text is explicitly excluded from the watermarking requirement.
2026 is only phase one. Starting in 2027, large platforms will have to preserve and expose the provenance data already present in the content they host — not just the AI providers that generate it. By 2028, the obligation extends to manufacturers of new recording devices, which will need to be able to embed this same provenance information.
This law is the regulatory reason behind moves like the Claude watermark we covered — although that Anthropic change was a direct response to the EU AI Act, the same “all synthetic content must carry a verifiable mark” logic now has the force of law in two of the world's largest markets (the EU and California) at nearly the same time, which pushes the technical standard (C2PA and similar) toward becoming the global minimum common denominator instead of a region-by-region implementation.
None of this law's obligations fall on someone who simply uses a generative AI tool to produce marketing or product content — they fall on the tool's provider. But for a company distributing that content, it's worth understanding that starting in 2026, a large share of visual content generated with the leading tools will carry detectable provenance metadata, which changes the risk calculus of passing off synthetic content as real — it keeps getting easier to verify, not harder.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel