California’s AI content-labeling law takes effect August 2 — the same day as the EU’s

California’s AI Transparency Act — the state’s requirement that widely used generative-AI systems make their own output detectable as machine-made — becomes operative on August 2, 2026, seven months later than lawmakers originally planned and, depending on which legal tracker is read, the same day the European Union’s own AI-content transparency rule takes effect. The law started as SB 942, signed by Governor Gavin Newsom in September 2024 with a January 1, 2026 start date. AB 853, signed thirteen months later, delayed that date to August 2, 2026 and added two more categories of company the original law didn’t reach.

What a covered provider must do starting tomorrow

Starting tomorrow, any “covered provider” — defined as anyone who creates a generative AI system with more than 1 million monthly visitors or users that’s publicly accessible in California — owes readers three separate things, not one. First, a free, public tool that lets anyone upload a piece of content or paste a URL and get back whether the provider’s own system made or altered it. Second, the option for a user to add a visible label to content they generate — clear, conspicuous and understandable, in the statute’s language — identifying it as AI-made. Third, and non-optional, a hidden machine-readable marker embedded in every image, video and audio file the system generates, carrying the provider’s name, the system’s name and version, and a timestamp — detectable by the provider’s own tool whether or not a user chose the visible label. A provider that licenses its system to a third party has 96 hours to cut off that licensee’s access if the licensee disables any of the three.

Who’s actually bound, and how the state can enforce it

The 1-million-user threshold is broad enough to catch essentially every foundation-model lab with a consumer-facing product — OpenAI, [Anthropic](#/company/anthropic), [Google](#/company/google) DeepMind, [Meta](#/company/meta) and [xAI](#/company/xai) all clear it on any plausible count of their chatbots’ monthly reach. The law exempts video games, television, movies and streaming, and any interactive experience built entirely from content users supplied themselves rather than the system’s own output. Enforcement runs through civil actions brought by the state attorney general or a city or county counsel, not a dedicated regulator — a prevailing plaintiff recovers attorney’s fees, and a court can issue an injunction against a licensee that keeps operating after its access was supposed to be cut off. Nothing in the statute requires a provider to announce compliance publicly, so whether a given company actually stood up a working detection tool by tomorrow is something the public will have to test for itself, one upload at a time.

The same date, an ocean apart

The European Union’s AI Act imposes its own transparency duty, Article 50, from the same August 2, 2026 date — providers of systems that generate synthetic audio, image, video or text must mark that output as machine-made and detectable, with a further grace period into December for systems already on the market before that date. That the two regimes converge on the same calendar day, one legislated in Sacramento and the other in Brussels, is the kind of detail that invites a tidy narrative: that California timed its law to match Europe’s. The legal trackers who cover this space don’t agree on whether that’s true.

What’s staged for 2027 and 2028

Tomorrow’s obligations aren’t the law’s last word. AB 853 layered in two more categories of covered entity on their own delayed schedule. From January 1, 2027, “large online platforms” — social media, file-sharing and messaging services and search engines with at least 2 million monthly users — must themselves detect and display the provenance data covered providers are now required to embed, and GenAI hosting platforms that let people download open model weights can’t knowingly distribute a system that lacks the required disclosures. From January 1, 2028, manufacturers of capture devices — phones, cameras, voice recorders — must offer a latent-disclosure option in newly produced hardware, extending the same provenance logic from the software that generates content to the devices that record it.

The precedent already in court

California has been here before, on a related law, and the result is worth reading before assuming tomorrow’s rollout goes smoothly. AB 2013, a separate 2024 statute requiring AI developers to disclose the categories of data used to train their systems, took effect January 1, 2026. OpenAI and Anthropic both published the required documentation — in each case, generalized categories rather than specific datasets, which is what the statute actually asks for. Google did not publish a matching disclosure by the same tracker’s count. Most consequentially, xAI sued California’s attorney general, Rob Bonta, in December 2025, arguing AB 2013’s disclosure mandate violates the First Amendment by compelling speech, the Fifth Amendment by taking trade secrets without compensation, and the Fourteenth Amendment through vague requirements. A federal district court denied xAI’s request to block the law while the case proceeds; xAI has since appealed to the Ninth Circuit.

None of that changes what happens tomorrow. Whether or not the shared August 2 date with Brussels was deliberate, and whether or not AB 853 eventually draws its own version of xAI’s lawsuit, the statute is now operative, the 1-million-user threshold now covers every major consumer AI product sold in the country’s largest state, and the $5,000-a-day penalty starts accruing the first day a covered provider is out of compliance. The open question isn’t whether the law exists. It’s whether anyone checks.