California's New AI Law Just Kicked In. Here's What Actually Changes

On August 2, California's AI Transparency Act became operative — the most detailed state-level rulebook in the country for labeling AI-generated content. It's easy to miss next to the daily churn of new model releases. It shouldn't be.

1. What the Law Actually Requires

The rules apply to "covered providers" — any generative AI system with more than 1 million monthly users in California, regardless of where the company is headquartered.

• They must offer a free, public tool that lets anyone check whether a piece of image, video, or audio content was created or altered by that provider's AI.
• They must attach both a visible label ("manifest disclosure") and embedded metadata ("latent disclosure," built on the C2PA provenance standard) to AI-generated content.
• Stripping that metadata — or building a tool designed to strip it — is explicitly prohibited.
• A follow-up law, AB 853, widens the net starting in 2027 to large online platforms (2 million+ users), AI-hosting platforms that distribute model weights, and even camera and device manufacturers, who'll need to embed authenticity data by default.

2. Why Now, and Why California

The original law was signed back in September 2024 and was supposed to take effect January 1, 2026. It got pushed to August 2 on purpose — timed to line up with the enforcement date of the EU's AI Act for high-risk systems, so major AI providers would face roughly the same disclosure clock on both sides of the Atlantic.

And because the threshold is simply "1 million monthly users in California," the law functions less like a state regulation and more like a de facto national one — any company big enough to have that many users there has to comply everywhere, not just in-state.

3. The Gap Between the Law and the Internet You'll Actually See

One state law now decides whether more than a million Californians can prove what's real.

That's the headline. Here's the catch: the detection tool only works for content made by a provider's own system — it can't retroactively label the flood of AI content already online, and it does nothing for tools that fall under the 1-million-user threshold. The bigger backstop — platforms actively detecting and surfacing provenance data on content they distribute, rather than just create — doesn't arrive until 2027.

So for now, the law changes what the largest AI providers have to disclose. It doesn't yet change what you'll actually see scrolling through an ordinary feed.

The Frame Feed Perspective

We've written before about crypto's slow absorption into the system it was meant to replace. This is close to the mirror image: after several years of AI content flooding the internet with effectively no way to verify its origin, the first real legal mandate to prove where something came from is arriving — not from Washington, not from Brussels, but from a single state's user-count threshold.

When "prove it's real" becomes a legal requirement instead of a courtesy, what happens to everything that still doesn't carry a label?

Related reads: Agentic Reality: Why 2026 is the Year AI Starts Working for You · Is Crypto Becoming Part of the System It Was Meant to Replace?