California's AI Transparency Act Takes Effect With Fines That Compound Daily

California's New AI Watermark Law Covers Deepfake Video and Voice — Not the Text Your Students and Employees Are Writing

California's AI Transparency Act — SB 942, as amended by AB 853 — took effect on August 2, 2026, after regulators moved the start date forward from a planned January 1, 2026 launch. It requires generative AI providers whose publicly accessible systems draw more than 1 million monthly visitors or users in California to offer a free public tool for checking whether a file came from their system, embed an invisible ("latent") watermark in generated media, and offer an optional visible disclosure. Providers named in coverage include OpenAI, Google, Meta, Midjourney, xAI, ElevenLabs and Suno. Violations carry a $5,000 civil penalty per day the violation continues, enforced by the California Attorney General or a city or county counsel. The law took effect the same day EU AI Act transparency duties under Article 50 began applying.

Does this law actually cover AI-written text?

No — the statute's disclosure and watermarking duties are written around images, video and audio, not plain text output.

Read the requirements closely and the pattern is consistent: a public verification tool for "image, video or audio," a latent disclosure embedded in "those files." Nothing in the coverage describes a parallel obligation for the text a chatbot writes. If you run a newsroom, school district or agency that cares about catching AI-generated prose, this law does not put pressure on OpenAI, Google or Meta to watermark the essays, articles or emails their models produce. That gap is worth noticing precisely because the law's headline — "AI Transparency Act" — invites readers to assume it's comprehensive.

Which vendors are actually on the hook?

Only providers whose generative system is publicly accessible and crosses roughly 1 million monthly California users or visitors.

That threshold is generous to smaller players. The AI humanizers, paraphrasing tools and niche image generators that schools and publishers wrestle with day to day sit well below 1 million monthly users and face no obligation under this law. The compliance burden lands on the largest media generators — exactly the companies already building provenance tooling for other reasons — while the long tail of tools most likely to be misused stays untouched.

What counts as "compliant," and who's checking?

A free verification tool plus a watermark meeting "widely accepted industry standards" — a phrase the statute doesn't pin to one format.

C2PA, the provenance standard backed by Adobe, Microsoft and the BBC, is the obvious reference point, but the law doesn't require it specifically. Enforcement runs through the Attorney General or local counsel — there's no independent lab certifying that a given watermark actually survives cropping, compression or re-upload before a violation is charged. That's the same weak spot this site keeps flagging in detector-accuracy claims: a vendor's self-reported compliance is not the same as a third party measuring it.

What should buyers ask their AI vendor now?

Ask explicitly whether disclosure tools extend to text, which watermark standard is used, and who verifies it independently.

If you're procuring content-detection or provenance tooling, don't assume California's law does that work for you. Ask whether any "AI Transparency Act compliance" claim from a vendor covers written content at all, or only the images and video the statute names. Given how this site has covered content governance elsewhere, including Wipro's Ivana Bartoletti on AI content oversight, the pattern holds: transparency laws are arriving piecemeal, medium by medium, and text is consistently the one lagging behind.

Frequently asked questions

When did the law take effect, and why was the date moved?

It took effect August 2, 2026. AB 853 pushed the start forward from an originally planned January 1, 2026 operative date.

How big is the penalty?

$5,000 per violation, with each day of non-compliance counted as a separate violation, enforced by the state Attorney General or city/county counsel.

Does this replace text detectors like GPTZero or Originality.ai?

No. The law targets image, video and audio disclosure from large providers; text-detection tools remain the only check on AI-written content for now.

Source: Startup Fortune.