They Built a Verifier for the Fake.
There’s Still No Verifier for You.
The first provisions of the California AI Transparency Act went live on Sunday. Large AI systems must now stamp the images, video, and audio they generate with machine-readable proof of origin, and hand the public a free tool that reads the stamp. California just built the infrastructure that verifies what’s fake. No statute requires its general-purpose equivalent for what’s real. The asymmetry is the condition.
As of last Sunday, the machines carry papers. California’s AI Transparency Act is now operative: any generative AI system with more than 1,000,000 monthly visitors or users, publicly accessible in the state, must embed a latent disclosure in every image, video, or audio file it produces. Not a vague label. A record, so far as technically feasible and reasonable: the provider’s name, the system and version that generated it, the time it was made, a unique identifier. And every covered provider must offer a public detection tool, “at no cost to the user,” that reads those papers back out. The penalty for skipping any of it is $5,000 per violation, with each day counted as a fresh one.
The act (SB 942, authored by state Sen. Josh Becker of Menlo Park) was signed in September 2024 and originally set to take effect on New Year’s Day. A follow-on bill, AB 853, pushed the operative date to August 2, a timing that contemporaneous legal analysis read as deliberate. It landed on the same Sunday the European Union’s AI Act transparency obligations became enforceable law, the statute Edition No. 031 read closely. The architecture extends forward from there: platform-level obligations arrive in 2027, and cameras and recording devices sold in California from 2028 must offer provenance at the point of capture. Read the sequence for what it is: two jurisdictions, two continents, one operative date, one direction. A single law pointing at the synthetic is a policy choice. Two arriving in step are an architecture.
Every obligation in both laws runs in the same direction: at what machines make and what machines are. The mandated metadata answers exactly one question with legal force: did a machine produce this? Type the harder question into any verifier you like: did this person actually earn the authority their record claims? Silence. That question has no covered provider. The authentic has no detection tool. The state has decided the public deserves certainty about what is fake. What is real is still left to inference.
The concession built into the statute
For years, the standing excuse for why the recognition layer runs on guesswork was a practical one: there is no infrastructure for verifying identity and origin at machine scale. That excuse died on Sunday. The legislature just mandated exactly that infrastructure (identity, origin, vintage, and a unique identifier, embedded and machine-readable, so far as technically feasible and reasonable) for every image, video, or audio file a large model emits. Feasibility is off the table; the only question left is direction. And notice what those four fields become when you read them as identity rather than forensics: who stands behind this, what body of work produced it, what vintage it carries, what ties it to everything else from the same source. Those are precisely the four fields most professional records have never carried in machine-readable form.
The fourth clinical characteristic of Digital Derangement Syndrome™, Trust Transfer Failure, describes real-world authority that never gets translated into machine-readable trust. The answer engines deciding who gets cited and recommended do not read diplomas, courtrooms, or a twenty-five-year client roster. They read signals, and where the signals are unstructured, they infer.
Untreated, the asymmetry compounds on a schedule. Each phase the provenance layer adds (platforms in 2027, capture devices in 2028) makes the machines’ side of the ledger more legible, and inference-by-contrast harsher on the side that never translated: the more of the world that carries verifiable origin, the more conspicuous a record that carries none. Trust Transfer Failure was never a static gap; each improvement on the machines’ side widens it.
Named for the record. California AI Transparency Act, SB 942 (2024), operative August 2, 2026, as amended by AB 853 (2025). Covered content is image, video, and audio; text output is outside the act’s disclosure scheme. Civil penalty: $5,000 per violation, enforced by the Attorney General, a city attorney, or a county counsel. Nothing in this edition characterizes any provider’s compliance status.
What this means if you never generate a pixel
You could run a practice for thirty years and never touch an image model, and this law still redraws your terrain. The machines’ output now arrives with papers. Your record still arrives without them. Every quote of yours circulating unattributed, every result summarized with your name sanded off, every credential living on a page the engines read as unstructured prose: all of it is exactly as unverifiable this week as it was before the act took effect. Sunday changed the baseline you are measured against, not the record you carry.
That is the actual assignment. Answer Engine Authority™ is, in substance, the record SB 942 will never mandate on your behalf, and four of its six phases answer the same questions the machines’ papers now answer. Entity architecture settles who stands behind the claim; signal consolidation gathers the work into one legible body. Third-party corroboration brings other voices to it, and ongoing signal maintenance keeps the record current. The state will make the machines confess what they made. Making yourself legible was never going to be legislated. It has to be installed.
California AI Transparency Act, SB 942 (Becker), signed September 19, 2024, codified in the Business and Professions Code beginning at §22757. The covered-provider definition (generative AI systems with over 1,000,000 monthly visitors or users, publicly accessible in California), the mandatory latent-disclosure fields, the optional manifest disclosure, the detection-tool requirement, and the $5,000-per-violation penalty are read directly from the enrolled text at leginfo.legislature.ca.gov. The quoted phrase “at no cost to the user” is §22757.2(a) verbatim.
Operative-date amendment: AB 853 (2025), delaying the act from January 1 to August 2, 2026 and phasing hosting-platform (2027) and capture-device (2028) obligations, corroborated by Morgan Lewis (August 2026) and Orrick (October 2025). The reading of the new date as deliberate EU AI Act alignment is Troutman’s (October 2025); AB 853’s own text does not state a rationale. Text-only AI content is outside the act’s disclosure scheme, and this edition makes no claim about any provider’s compliance. Amendment activity continues: SB 1000 (Becker), a pending urgency bill, would further revise these provisions, including deleting the user threshold; this edition describes the act as it stands at publication.