California's AI Transparency Act became operative on August 2, and for the first time the state now requires the largest generative AI systems to label what their machines produce. The law reaches any publicly accessible service with more than one million monthly users, which pulls in most of the tools people actually use.
The core of it is disclosure. Covered providers have to offer a free tool that lets anyone check whether a piece of content came from their AI. They must embed a hidden, machine readable signal in the images, video, and audio their systems generate, and they have to give users the option to attach a visible label as well. Those obligations also flow down to companies that license the underlying models, so a provider cannot escape the rules by selling access to someone else. The statutory text sets a civil penalty of 5,000 dollars per violation, and each day out of compliance counts as a separate one.
Why the date moved
The Act was originally written to start on the first day of 2026. A follow up bill, AB 853, pushed the effective date to August 2 so that California would line up with the European Union, whose own transparency rules for general purpose AI switched on the same week. Legal teams at firms including Morgan Lewis spent the summer telling clients to treat the two regimes as one compliance project rather than two. Our own coverage of the EU AI Act deadline traced how close those calendars now run.
A law that may change under it
There is an unusual wrinkle. A separate bill, SB 1000, would rewrite parts of the Act even as it takes effect, potentially removing the one million user threshold and dropping the visible label requirement. Urgency bills move quickly in Sacramento, so the version of the law that companies scrambled to meet this month could look different by the autumn. For now the compliance teams are building to the text on the books.
The wider drift
Watermarking and detection have moved from voluntary pledges to statute in the space of a year, and California is not alone. The same instinct sits behind Minnesota's fight over synthetic nudes, which we reported last week, and behind the EU's machine readable marking rules. Whether a hidden signal survives a screenshot or a re-encode is the question engineers keep raising, and the honest answer is that today's methods are not tamper proof. The law is betting that a shaky signal still beats none at all.
Commentarii · 0