California Now Makes AI Labs Ship a Detector for Their Own Output

TL;DR

As of today, August 2, the California AI Transparency Act (SB 942) is operative. If your generative system has more than 1,000,000 monthly visitors or users and is publicly reachable in California, you now owe three things: machine-readable provenance embedded in every generated image, video, and audio clip; an option for users to stamp a visible "AI-generated" label; and a free, publicly accessible detection tool, with an API, that identifies content your own system produced. Civil penalties run $5,000 per violation, with each day counting as a separate violation. The nastier clauses do not land on model builders at all: from January 1, 2027, sites that distribute model weights inherit a duty of their own.

What actually switched on today

SB 942 passed in 2024 with a January 1, 2026 start date. Then AB 853, signed October 13, 2025, rewrote it: new obligations, new categories of defendant, and a new operative date of August 2, 2026, chosen to land on the same day as Article 50 of the EU AI Act. Sacramento moved its own deadline seven months so it would sync with Brussels, which is the closest thing to interoperability that regulators have ever shipped.

The trigger, at section 22757.1(d), is a person that "creates, codes, or otherwise produces a generative artificial intelligence system that has over 1,000,000 monthly visitors or users." Note what that counts: the system, not the company. A three-person team whose image generator crossed a million monthly users is covered. A large enterprise whose internal tool serves 4,000 employees is not.

Three duties, and one of them is weird

Section 22757.3 scopes the content duties to AI-generated "image, video, or audio content, or content that is any combination thereof." Text is out. Your LLM wrapper is not the target here; your diffusion endpoint is.

1. Latent disclosure

Every generated image, video, or audio file carries machine-readable provenance recording the provider's name, the system's name and version, the time and date the content was created, and a unique identifier. It has to be permanent or extraordinarily difficult to remove, to the extent that is technically feasible. In practice that means C2PA Content Credentials, optionally paired with an imperceptible watermark like Google DeepMind's SynthID, because metadata alone dies the first time someone screenshots your output.

2. Manifest disclosure

Users get the option to attach a clear and conspicuous visible label. This is the mildest of the three and the one most products already ship.

3. The detection tool

Here is the unusual one. You must publish a free tool, open to the public, that accepts an upload or a URL and reports whether the content came from your system, returning the provenance data it finds. It has to expose an API, not just a web form. It cannot retain what people submit beyond producing a result, and it cannot harvest personal information from the people using it.

Read that again as an engineer, because it is a service, not a checkbox. California is requiring you to operate public infrastructure whose entire purpose is to identify your own output, on your dime, at whatever traffic the internet decides to send it. It is a bit like requiring a distillery to run the free breathalyzer in the parking lot, keep it calibrated, and never ask who is blowing into it.

The 96-hour kill switch

If you license your system to third parties, the contract has to require them to keep the latent disclosure intact. If you find out a licensee has modified the system to strip it, you have 96 hours to revoke their access, and they must stop using it once you do. Four days, from discovery to termination, on a clock you do not control.

Anyone who has watched an enterprise contract-termination workflow move will recognize the problem. The engineering answer is to make revocation a runtime capability rather than a legal one: key rotation, per-licensee tokens, and a switch someone on call can actually flip.

January 2027 is the clause open-weight people should read

The provider duties are the headline, but AB 853 added two categories that switch on January 1, 2027, and one of them points straight at model distribution.

• Large online platforms (section 22757.3.1), defined as exceeding 2,000,000 unique monthly users over the preceding 12 months, must detect provenance data in content they distribute and surface it to users, so people can see whether something was generated, altered, or captured by a device.
• GenAI hosting platforms (section 22757.3.2), defined as a site or application that makes available for download the source code or model weights of a GenAI system to a California resident, "shall not knowingly make available a GenAI system that does not place disclosures."

That second one is aimed at exactly the surface you use every day. A site like Hugging Face distributes weights; the weights, once downloaded, run on someone else's machine, in someone else's harness, with no obligation of their own to sign anything. A downloaded checkpoint has no idea it is supposed to be honest about itself. How a hosting platform is meant to know, at upload time, whether a set of weights will "place disclosures" downstream is the open question, and the word doing the load-bearing work is knowingly. Five months is not a lot of runway to find out what a court thinks that means.

What non-compliance costs

Section 22757.4(a)(1) sets the penalty at $5,000 per violation, and each day of noncompliance is its own violation. There is no cap in the arithmetic, and prevailing plaintiffs collect attorney's fees and costs. Enforcement runs through the Attorney General, city attorneys, and county counsel.

How it differs from the EU rule landing the same day

Both regimes want machine-readable marks on synthetic media, and both start today. The differences are the part worth internalizing.

• Threshold. The EU rule applies regardless of size. California only bites above 1,000,000 monthly users on the system.
• Scope. Article 50 covers text, chatbot disclosure, and deepfake labeling. California's latent disclosure covers image, video, and audio, and leaves text alone.
• The detector. Brussels asks you to mark your output. Sacramento asks you to mark it and operate the tool that finds it. There is no EU analogue to that duty.
• Penalty shape. The EU tops out at a percentage of global turnover. California charges by the day, which is smaller per unit and considerably more patient.

If you are near the line

Pull your actual monthly unique users for anything that generates media, not your marketing number. If you are anywhere near seven figures, the cheap move is to turn on C2PA signing now, since your model host probably already exposes it, and treat the detection endpoint as a real service with a rate limit, an SLO, and an owner. If you license the system to anyone, add the disclosure-preservation clause and build the revocation path before you need it at hour 90 of 96.

Key Takeaways

• California's AI Transparency Act (SB 942, as amended by AB 853) became operative August 2, 2026, deliberately aligned with the EU AI Act's Article 50 date.
• Covered providers are systems with more than 1,000,000 monthly visitors or users that are publicly accessible in California; the threshold attaches to the system, not the company.
• Three duties: latent provenance disclosure in generated image, video, and audio; an optional visible label for users; and a free public detection tool with an API that identifies your own output.
• Penalties are $5,000 per violation with each day a separate violation, enforced by the AG, city attorneys, and county counsel, with attorney's fees to prevailing plaintiffs.
• Licensee stripping the disclosure means a 96-hour window to revoke access, so make revocation a runtime capability, not a legal one.
• From January 1, 2027, platforms above 2,000,000 monthly users must surface provenance, and sites distributing model weights may not knowingly host GenAI systems that fail to place disclosures. Capture devices follow on January 1, 2028.

Sources: California SB 942, California AI Transparency Act, California AB 853 bill text, Troutman Pepper Locke: California AI Transparency Act Amendments Signed Into Law, AI Laws by State: SB 942 compliance guide, National Law Review: California Establishes AI Transparency Act, Crowell & Moring: CAITA amendments