California GenAI Law: $5,000 Daily Fines When AI Video Hides Origin

Headlines that say California will fine you unless AI video has a visible watermark are directionally right but legally imprecise. The state's California AI Transparency Act, enacted as Senate Bill 942 in 2024 and amended by Assembly Bill 853 in 2025, does not only mandate a sticker on the frame. It requires latent disclosures embedded in AI-generated or AI-altered image, video, and audio, plus a free detection tool, for covered providers above a million-user threshold, with civil penalties of five thousand dollars per violation and each calendar day of noncompliance treated as a separate violation. At that rate, sustained noncompliance can reach seven figures annually without any need for creative statutory reading.

Primary sources: SB 942 chaptered text (September 19, 2024) and AB 853 chaptered text (October 13, 2025), both codified in Division 8 of the California Business and Professions Code starting at Section 22757. This piece translates those sections for builders who ship generative models, APIs, or consumer apps that output synthetic media.

On March 30, 2026, Governor Gavin Newsom also signed an executive order framed as strengthening AI protections and responsible use, including direction for state agencies on watermarking AI-generated or heavily manipulated images and video used in government contexts (California Governor news release). That order is separate from the private-sector duties in SB 942 and AB 853, but it signals the same policy vector: treat synthetic media as something the public should be able to trace, not guess.

What the law requires beyond a visible watermark

A covered provider must do two different things for image, video, or audio (or combinations) that its GenAI system creates or alters.

First, it must offer users an option for a manifest disclosure: something a reasonable person can perceive, clear and conspicuous, appropriate to the medium, and permanent or extraordinarily difficult to remove where technically feasible. That is the closest statutory cousin to what people colloquially call a watermark.

Second, and easy to miss in social summaries, the provider must include a latent disclosure in AI-generated image, video, or audio created by the system. Latent is defined as present but not manifest. The latent payload must to the extent technically feasible and reasonable convey the provider name, GenAI system name and version, timestamp, and a unique identifier, either directly or through a link to a permanent website. It must be detectable by the provider's own AI detection tool, consistent with widely accepted industry standards, and again permanent or extraordinarily difficult to remove where feasible.

So the compliance surface is provenance metadata plus UX disclosure plus a public detector, not a single PNG stamp.

Who is actually regulated

The statute defines covered provider as a person that creates, codes, or otherwise produces a generative artificial intelligence system with more than one million monthly visitors or users that is publicly accessible within California. That is a scale gate, not a rule for every indie LoRA on Hugging Face.

AB 853 extends the penalty section so each day that a covered provider, large online platform, or capture device manufacturer is in violation is a discrete violation. It adds large online platform duties (provenance detection and UI starting January 1, 2027 for platforms over two million unique monthly users in the prior twelve months), GenAI hosting platform rules (no knowingly hosting models that skip required disclosures from January 1, 2027), and capture device latent disclosure defaults for devices first sold in California on or after January 1, 2028 (operative January 1, 2028). If you run infrastructure or hardware, calendar those later dates separately from the core GenAI provider obligations.

Exempted categories include products that provide exclusively non-user-generated video game, television, streaming, movie, or interactive experiences, per Section 22757.5 as enacted in SB 942. Read the full carve-out before assuming your OTT or game pipeline is out of scope.

When obligations go live

SB 942 originally set the chapter operative on January 1, 2026. AB 853 delayed that to August 2, 2026 (Section 22757.6 as amended). If your roadmap still says New Year's Day 2026, update it: August 2, 2026 is the operative date for the core Transparency Act chapter unless a court says otherwise.

Platform and device add-ons follow 2027 and 2028 as quoted above. Legal teams sometimes file this under "AI governance program" rather than "watermark ticket," because the free detection tool requirements (upload or URL, API access, limits on retaining personal provenance data, feedback loop) are substantial engineering work on their own.

Why people say millions in fines

Section 22757.4 sets civil penalties at five thousand dollars ($5,000) per violation, recoverable by the Attorney General, a city attorney, or county counsel, with reasonable attorney's fees and costs for a prevailing plaintiff. The statute then states explicitly that each day of violation by a covered provider (and, after AB 853, the other regulated categories) is a separate violation.

Arithmetic is blunt. $5,000 × 365 days ≈ $1.83 million for a hypothetical year-long continuous violation by one entity, before fees. Multi-product or multi-jurisdictional enforcement narratives can scale from there. No, the text does not use the word million; yes, the per-day multiplier is how regulators and plaintiffs' counsel will model exposure. This is civil enforcement, not a per-download criminal fine, but it is enough to change how general counsel prices shipping an unmarked video API in California.

Third-party licensees get a different track: if they modify a licensed GenAI system so it cannot include required disclosures, the licensor must revoke within ninety-six hours of discovery, and the licensee must cease use after revocation. Violations of the cease-use rule can yield injunctive relief and fees, separate from the per-day penalty framing for the core provider duties.

How Newsom's March 2026 order fits

The March 30, 2026 executive order is chiefly about state government use of AI, contractor attestations, and best practices for watermarking consistent with state law. It does not replace or rewrite SB 942, but it amplifies the same narrative for vendors who sell to Sacramento: provenance and labeling are now part of the political brand of California AI policy, not an obscure appendix in the Business and Professions Code.

For developers, the practical split is simple. Selling synthetic media tools to consumers or enterprises in California pulls you toward B&P 22757 compliance engineering. Chasing state contracts adds procurement questionnaires about bias, CSAM safeguards, and watermarking expectations on top.

What product and infra teams should implement

Treat compliance as metadata plus UX plus verification, aligned with industry specs the statute already points to (widely accepted standards for provenance data and system provenance data).

Pipeline design: emit system provenance at generation time, bind it so re-encode strips are detectable as tampering where your threat model requires it, and wire the manifest path as a user-toggle that survives normal export paths for video.

Detector API: you owe users a free tool with public access (subject to reasonable security throttling), URL or upload input, and a machine API so third parties can batch-check assets. You cannot retain personal provenance from submissions, and you must minimize retention of uploaded content.

Licensing: if you OEM your model, contractual clauses must preserve disclosure capability, and your license revocation playbooks need a ninety-six hour clock.

For background on synthetic media abuse vectors and detection limits, see the deepfake and synthetic media developer guide. For model choice and API economics while you redesign outputs, pair this with the ChatGPT vs Claude vs Gemini vs Grok comparison and the LLM API pricing tracker.

What this does not solve

The Transparency Act targets labelling and provenance for large commercial GenAI systems. It does not stop open-weights models trained elsewhere from generating unmarked video on a laptop in Fresno. It does not automatically fix political deepfakes at scale, though it gives enforcement hooks against major distributors who ignore latent metadata. It also cannot eliminate adversarial stripping of disclosures; the statute already caps duties at technical feasibility.

Expect standards forums (C2PA-style provenance, watermarking research, codec-level signalling) to keep moving faster than any one state bill. California just attached real money to the provider-side half of the stack for the biggest platforms.

Key Takeaways

• SB 942 (2024) created the California AI Transparency Act; AB 853 (2025) delayed the operative date to August 2, 2026, and added 2027 platform and 2028 capture-device rules (legislative text).
• Covered providers are builders of GenAI systems with >1,000,000 monthly visitors or users publicly accessible in California; they must ship latent and optional manifest disclosures on AI-generated image, video, and audio, plus a free detection tool with an API.
• Civil penalties are $5,000 per violation, with each day of violation counting as separate; a full year of continuous breach implies about $1.83M in penalties before attorney fees.
• Large online platforms (>2M unique monthly users) face provenance detection and UI duties from January 1, 2027; GenAI hosting platforms must not knowingly distribute models that skip disclosures from the same date.
• Governor Newsom's March 30, 2026 executive order pushes state agency watermarking and contractor guardrails in parallel with the statutory scheme (Governor's office).
• Engineering implication: treat this as provenance metadata + user-visible labelling + verifier services, not a single visible watermark checkbox.