AI

California’s deepfake law goes live Sunday. Text was quietly cut out of it.

The AI Transparency Act becomes operative August 2 after a seven-month delay. Any AI system with a million monthly users has to ship a free detection tool and embed provenance data in what it makes — images, video and audio only. The penalty is $5,000, and every day counts as a fresh violation.

N Noah · The Sharp Brief · July 31, 2026 · 4 min read

On Sunday the California AI Transparency Act becomes operative, and the first state-level content-provenance mandate in the United States stops being theoretical. It was supposed to bind on January 1. AB 853, signed by Governor Gavin Newsom on October 13, 2025, moved the date to August 2, 2026 — not to give anyone a break, but to line California up with the EU AI Act rather than with Washington.

It reaches “covered providers”: anyone producing a generative AI system with more than one million monthly visitors or users that is publicly accessible in California. Four obligations attach. Publish a free, public detection tool that lets anyone check whether a piece of content came out of your system. Offer users a manifest disclosure — a visible label on the output. Embed a latent disclosure, machine-readable provenance carried inside the file itself. And write terms into your licensing agreements so downstream licensees cannot strip any of it out.

Then read the scope line carefully. The statute’s definition of a generative AI system mentions text. The word was removed from the operative provisions. As Troutman Pepper Locke noted in its analysis of the amendments, the obligations apply only to image, video and audio. The most common thing people do with AI — generate written words — sits entirely outside a law written to fight deception.

Our take: Price the enforcement before you price the compliance. The penalty is $5,000 per violation, collectible by the Attorney General, any city attorney, or any county counsel — and each day out of compliance is a discrete violation. A year of quiet non-compliance on one obligation clears $1.8 million, and a prevailing plaintiff collects fees on top. Three sets of lawyers with concurrent authority and a per-day clock is a structure built to be used. But the text carve-out tells you what this actually is: not an AI honesty law, a photo-and-video provenance law. It will make synthetic images traceable and do nothing about the synthetic paragraph in your inbox. The compliance question that matters is not whether you are covered on Sunday. It is whether you are covered on January 1.

The dates after Sunday

What to watch

California keeps writing the rules Washington has not, and keeps timing them to Brussels — the same pull that shapes Europe’s compute agenda. If you ship AI-generated media, Sunday is a hard date. For everyone else it is the last quiet weekend before proving what is real becomes the publisher’s job, not the reader’s. Until the tooling catches up, verifying it yourself is the only control that actually ships.

Advertisement

Get the day, decoded — at 7 PM ET

The Sharp Brief: AI, money, business & performance in five sharp minutes. Free.

Free bonus: subscribe today and The 2026 AI Playbook (PDF) lands with your welcome email.

Recommended by 5+ newsletters across AI, markets & business.