Something changed on your feed this weekend, even if you can't see it yet. As of August 2, the California AI Transparency Act — CAITA — is in effect, and the companies behind the big image and video generators are now legally on the hook for labeling what their systems create. They also owe the public a free tool for checking what's real.
The law that finally arrived
CAITA started as SB 942, which Gov. Gavin Newsom signed back in September 2024 with deepfakes squarely in mind. It was supposed to kick in on January 1 this year. Then AB 853 came along, moved the date to August 2, and widened the net — partly to line California up with the EU AI Act's transparency enforcement timeline. So the delay wasn't a retreat; it was a regroup.
Here's who's covered: anyone who makes a generative AI system with over a million monthly users that's publicly accessible in California. That's basically every major image and video generator you can name. One boundary worth knowing — the law applies to AI images, video, and audio. Text isn't covered.
Four requirements went live this weekend. Covered providers have to offer a free public detection tool, so anyone can check whether content came from their system. They have to let users add a visible "this is AI" stamp. They have to embed hidden, machine-readable provenance data in everything their systems generate. And they have to force the same rules on companies licensing their tech — with 96 hours to pull the license if a licensee strips the labels.
The fine sounds modest at $5,000 per violation. The catch: every day of noncompliance counts as a fresh violation. It's built to snowball.
Your feed gets labels next
August 2 was phase one. Phase two, on January 1, 2027, is the one you'll actually notice. That's when large online platforms — social media, file-sharing, and messaging services, plus search engines with over two million unique monthly users — must start detecting provenance data in what they distribute, show you whether something was AI-generated, and let you inspect the provenance trail yourself.
In other words, AI labels on posts stop being a voluntary experiment and become law in the largest state in the country. Sites that host downloadable models get their own 2027 requirements too.
Then comes the 2028 twist: cameras, phones, and voice recorders sold in California will have to support embedded-by-default disclosures recording which device captured something and when. The system is meant to work from both ends — real content gets a verifiable origin, synthetic content gets a flag.
Why it won't stay a California story
No AI company is going to ship a special California edition of its product. Once the labeling plumbing exists, it exists everywhere — which makes CAITA a de facto national standard, the same way the state's privacy law was. Attorneys at Bradley Arant Boult Cummings, writing in the National Law Review, have been warning clients for weeks to map their products against each deadline and fix their licensing contracts, calling the required engineering work substantial.
What you get out of it, starting now: official detection tools for interrogating a suspicious image. What you get next year: labels in the feed itself. Will provenance data actually stop deepfakes? That's genuinely unsettled — determined bad actors will keep finding workarounds. But the era of AI content moving through the world with no paper trail at all ended on Sunday.
Image: Sanket Mishra, via Pexels
