California's AI Transparency Act became operative on 2 August
Generative AI providers serving California must now offer watermarking, latent disclosures and detection tools for AI-generated content. It is in force now — unlike the EU obligations it resembles, which moved to December 2027.
The comparison writes itself. Europe legislated first, comprehensively, and its high-risk obligations now arrive in December 2027. California legislated narrowly, on one property of one category of system, and it has been operative since 2 August.
The obligations are specific enough to build against: watermarking of generated content, latent disclosures that survive ordinary handling, and detection tools made available so third parties can check provenance. That last one is the demanding part — publishing a detector is a commitment to something falsifiable.
Narrow scope is why it shipped. There is no conformity assessment, no notified bodies, no harmonised standards to write first. Provenance is a property you can specify without deciding what counts as high risk, which is the question that took Europe an extra eighteen months.
It also lands where the market was already going. Google shipped SynthID on by default before any of this bound, and a platform that can already demonstrate which outputs were generated is holding the compliance artefact. Firms that made provenance optional now have a deadline that has already passed.
Vorp Labs — US AI Regulation Update: July 2026 Laws & Policy → · Gunderson Dettmer — 2026 AI Laws Update: Key Regulations and Practical Guidance → · Baker Botts — U.S. Artificial Intelligence Law Update →