California’s AI Regulation Moves from One Flagship Law to a Full Deployment-Layer Stack — and It All Lands on Newsom’s Desk

On the sine die night of August 31, the California Legislature sent Governor Newsom not one AI bill but a portfolio — shifting AI governance from the model layer to the deployment layer in a single session, with every measure now hostage to one signature on a fixed clock.

What Happened

The California Legislature’s official bill-history record — leginfo.legislature.ca.gov — shows a legislature that spent its final session night sending Governor Gavin Newsom a portfolio of AI measures spanning nearly every surface where an AI system touches a person. A critical preliminary: several of the August 31 floor votes were taken in the closing hours of sine die, and the official bill-history system posts actions with a lag. Vote counts and enrollment status on the newest measures are late-night snapshots, not certified records. More critically, two of the highest-reach bills — SB 1119 (Padilla), on companion chatbots embedded in products aimed at minors, and AB 2564 (Ward), the broad ban on AI-driven “surveillance pricing” — had not posted confirmed final concurrence votes at the deadline. Neither should be reported as passed until those votes appear in the official record. And passing the Legislature is not the same as becoming law: every one of these bills can still be vetoed before September 30. Newsom has vetoed high-profile AI legislation before.

With those constraints explicit, what the record does confirm as enrolled or presented to the Governor is substantial. SB 813 (McNerney) would create a state AI Standards and Safety Commission and a framework for independent, third-party assessment of AI models — the Senate concurred 37–0 on August 30. AB 1883 (Bryan) would make California the first US state to bar employers from using AI to infer workers’ emotional states or harvest neural data. SB 947 (McNerney) and SB 951 (Reyes) address automated decision systems in hiring and AI-driven job-displacement notice, respectively, with floor concurrence votes on the night of August 31. SB 503 (Weber Pierson), targeting AI bias in health-coverage decisions, was presented to the Governor on August 30. AB 1979 covers AI in health care more broadly. SB 903 (Padilla), which limits AI from substituting for licensed therapists, passed 40–0. SB 867 (Padilla) bars companion chatbots embedded in children’s toys. AB 1609 (Zbur) requires disclosure when a customer-service chatbot is not a human. SB 1111 (Ashby) governs unauthorized digital replicas of real people. SB 1000 (Becker) — amendments to the AI Transparency Act — carries urgency status and would take effect on signature, not January 1, 2027. AB 2713 addresses AI provenance.

Everything else signed by the Governor — absent urgency clauses — takes effect January 1, 2027. The 2025 session’s SB 53, the frontier-compute disclosure law that set the prior legislative baseline, still binds. This week’s stack sits on top of it, not in place of it.

The Structural Read

In 2025, the California AI debate was organized around one question — the frontier — and one instrument: SB 53, a compute-threshold disclosure law aimed at the most powerful models during training. That was a natural first move. The frontier was where the novelty and the fear concentrated, and a disclosure rule for the biggest labs was the minimum legible intervention. What arrived on Newsom’s desk this week is structurally different. The 2026 portfolio is not about how AI systems are built. It is about what they are pointed at: hiring and termination (SB 947, SB 951), health coverage and mental-health therapy (SB 503, AB 1979, SB 903), children’s products (SB 867, and pending SB 1119), workers’ inner states (AB 1883), digital likenesses (SB 1111), customer-service disclosure (AB 1609), and AI-model transparency and provenance (SB 1000, AB 2713). This is regulation moving downstream — from the model layer to the deployment layer, from “how you train it” to “what you do with it once it ships.”

That migration is the maturation signal for AI governance. Compute-threshold rules were the right first instrument because they targeted the place of highest novelty. But AI’s actual economic and social footprint is not at the training run — it is at the application surface. Hiring systems, insurance algorithms, therapy substitutes, companion products, pricing engines: these are where AI meets people at scale, and they are precisely the layer the 2026 session chose to address. The direction of travel is clear even before a single bill is signed.

SB 813 is the connective tissue that makes the stack architecturally coherent. A compute-threshold disclosure law like SB 53 could require attestation without specifying who evaluates it. SB 813’s answer — a state AI Standards and Safety Commission with a framework for independent, third-party assessment — is the institutional “who actually checks” layer SB 53 deliberately left out. It is a move from self-attestation toward something closer to certification infrastructure. Whether that infrastructure is well-designed, adequately funded, and insulated from regulatory capture are questions that survive the signing ceremony. But the structural logic — disclosure requires an evaluator; SB 813 creates one — is sound.

The “California as de facto national regulator” pattern is not a boast — it is a structural observation from the privacy and emissions fights. California’s market size means that companies optimizing for a single compliance surface choose the California standard, even in states with no equivalent law. That dynamic does not require Washington to do nothing; it operates alongside federal inaction and survives it. The 2027 compliance surface for AI deployment — what your hiring algorithm must disclose, what your health-coverage tool must not infer, what your customer-service bot must say upfront — is being drawn in Sacramento this week, not in Congress. That is the pattern. It is not a guarantee the full stack survives intact, and it is not a substitute for federal action. It is the observed default.

Bill Status at Deadline