California AI Transparency Act: watermarking and detection duties take effect

As of 2 August 2026 the California AI Transparency Act (CAITA) is operative. SB 942, amended by AB 853 and signed last October, defines a «covered provider» as any entity offering a generative artificial intelligence system accessible in the state with more than one million monthly visitors or users. On these entities the legislature imposes a precise package of technical measures: a free public detection tool — anyone can upload a file or paste a URL and find out whether that content came out of that system, including through an API, without the provider being allowed to retain the personal data of whoever queries it — an option for the user to attach a visible label (manifest disclosure) and, above all, a permanent «latent disclosure» embedded in generated content. This digital watermark must carry the provider's name, the system version, the date and time of creation and a unique identifier, making the file technically traceable, and it must be detectable by the provider's own verification tool.

The penalty machinery set out in the text is rigid. Every day of non-compliance counts as a separate violation, punished with a civil penalty of 5,000 dollars. Enforcement lies with the California Attorney General, city attorneys or county counsel, and the prevailing party can recover its legal costs. Liability extends contractually down the distribution chain: if a third-party licensee alters the system to strip out the disclosures, the provider must revoke the licence within 96 hours. It is an approach that tries to lock down the provenance of content upstream, handing creators control over how their technologies are used downstream.

The statute lays out a calendar of progressive expansion. If 2 August marks the start for providers themselves, from 1 January 2027 the duties to detect and retain provenance data will fall on «large online platforms» with more than two million unique monthly users and on hosting platforms, which will not be allowed to knowingly distribute systems lacking the required labels. A year later, in 2028, the burden shifts to makers of capture devices, which will have to offer the option of embedding the latent disclosure, switched on by default. Where Europe, with article 50 of the AI Act, requires that the reader be told, California also requires the technical infrastructure to check.

There is no shortage of open questions, though. There are currently no verifiable figures on how far providers have actually adopted the new specifications. No enforcement action by the Attorney General has surfaced so far. One digital policy tracker describes the law as taking effect alongside a grace period running to 1 January 2027, a reading that does not match the law firms consulted, for whom 2 August 2026 is the full operative date for covered providers while 1 January 2027 concerns the new categories introduced by AB 853. The phrase does not appear in the statute. The scope of the law could also be redrawn by SB 1000, currently before the Senate, which would scrap the one-million-user threshold and change the nature of the verification tools. Finally, the technical problem of open-weight models remains open: the ability to modify weights downstream could make the guarantees the law imposes at the source ineffective.

— Olya