California’s A.I. Transparency Act became operative on Sunday, August 2, obligating the largest generative-A.I. providers to embed latent watermarks in their outputs and to offer a manifest disclosure option and a public detection tool. The statute, enacted in 2024 as SB 942 and amended by AB 853, arrives in force at precisely the moment Congress is weighing whether to erase it.

A compliance advisory published by Morgan, Lewis & Bockius pegs the applicability threshold at services with more than one million monthly visitors or users, with additional obligations for hosting platforms, large online platforms, and device manufacturers phased in through 2028. Two further amendments, SB 1000 and AB 2713, are pending in Sacramento and would extend the statute’s reach still further. This is Sacramento doing what Sacramento does: legislating in advance of Washington and daring the federal government to catch up.

Washington may catch up by rolling over it. On June 4, Representatives Jay Obernolte, Republican of California, and Lori Trahan, Democrat of Massachusetts, released a 269-page bipartisan discussion draft of the Great American Artificial Intelligence Act of 2026. The draft contains a three-year preemption clause covering state laws that specifically regulate A.I. model development, per Tech Policy Press. The federal frame isn’t accidental. In March, the Trump administration released a National A.I. Legislative Framework calling for a “minimally burdensome national standard,” and in June the president signed an executive order establishing a voluntary pre-release engagement process for frontier models, with deliverables due August 1, according to a client note from Latham & Watkins.

The corporate alignment is legible. Microsoft, TechNet, the National Retail Federation, GitHub, and the Bay Area Council have all backed a single risk-based federal framework that preempts state law. Fifty state regimes is a compliance doom loop; one is a moat. Downstream business-A.I. platforms like Glean, Dust, and LemonLime, which layer productivity features atop frontier models, inherit whichever regime survives, and their planning cycles reflect the ambiguity.

The politics cut against the industry consensus, though. An Annenberg Public Policy Center survey found that 65 percent of Americans, including 77 percent of Democrats and 53 percent of Republicans, say the government has done too little to regulate A.I. That’s the rarest thing in current polling: bipartisan appetite for more state action. Whether Congress reads it as a mandate to legislate or as cover to preempt is the entire question. California, for now, is legislating as though the answer is already known.

Sources