California's New AI Transparency Law, Explained Simply
As of August 2, 2026, California's AI Transparency Act (CAITA) requires major AI image, video, and audio tools to offer a free public detection tool and label the content they generate. Because most big AI companies operate nationwide, this California-specific law is already reshaping how AI tools work for people across the US. Here's what actually changed.
⚡ Quick facts
- Operative since: August 2, 2026 (delayed from an earlier planned date)
- Who it covers: "Covered providers" — generative AI systems with 1M+ monthly visitors/users, publicly accessible in California
- What's regulated: AI-generated image, video, and audio content (not text)
- Penalty: $5,000 per violation, per day — a 30-day lapse can mean $150,000 in liability
What CAITA actually requires
Companies that qualify as "covered providers" must now: offer a free, public AI detection tool anyone can use to check if content came from their system; give users the option to add a manifest disclosure — a visible label or watermark — to AI-generated content; embed a latent disclosure, an invisible signal baked into the file itself, so the origin can be verified even if a visible label is removed; and pass these same obligations down to any licensees of their AI system.
Who actually has to comply
The law defines a "covered provider" as anyone who creates or produces a generative AI system with more than one million monthly visitors or users that's publicly accessible within California. In practice, that threshold catches nearly every major AI image, video, and audio generator on the market — Google, OpenAI, Meta, Midjourney, and others all clear that bar many times over.
Text is not covered
It's worth being precise here: CAITA's disclosure requirements apply to image, video, and audio content — not text. A chatbot's written response isn't required to carry any of these disclosures under this law, even though the underlying model is generative AI.
What's coming next
CAITA is being rolled out in phases. The core obligations above became operative on August 2, 2026. Amendments signed into law (AB 853) extend additional requirements to large online platforms and GenAI hosting platforms starting January 1, 2027, and to certain capture device manufacturers starting January 1, 2028 — meaning the law's reach is still expanding.
Why this matters even outside California
Legally, CAITA only binds systems accessible within California. Practically, building a separate California-only version of a mainstream AI tool is rarely worth the engineering effort, so companies tend to ship the same detection tools and labeling features to everyone. This is the same dynamic this site covered with the EU's AI transparency rules and WhatsApp's new AI content labels — a state or regional law effectively becomes a global product feature.
C2PA Metadata Standards & Watermark Robustness Analysis
California's AI Transparency Act (CAITA) mandates that synthetic media generators embed C2PA (Coalition for Content Provenance and Authenticity) binding credentials directly into image and video file headers. These cryptographic provenance records chain creator identity assertions from camera sensor fingerprints through editing software signatures to final publishing metadata, making it computationally verifiable whether content was AI-synthesized or originally captured by human photographers.
Frequently asked questions
What is the California AI Transparency Act?
A California law, operative since August 2, 2026, requiring major generative AI providers to offer a detection tool and label AI-generated image, video, and audio content.
Who has to comply with CAITA?
"Covered providers" -- generative AI systems with over 1 million monthly visitors or users, publicly accessible in California.
What kind of AI content does it cover?
Image, video, and audio -- text generated by AI is not covered by CAITA's disclosure rules.
What happens if a company doesn't comply?
$5,000 per violation, with each day of non-compliance counted separately -- up to $150,000 for a 30-day lapse.
Does this apply outside California?
Legally it's California-specific, but since most AI tools operate nationwide, the labeling features tend to reach users everywhere in practice.