Thirty AI Bills Just Survived in Sacramento. The Next Four Weeks Set the US Floor.

Friday was the crossover deadline in the California Legislature, the procedural gate where a bill has to pass its first chamber or sit out the rest of the year. Nearly all of the state's roughly 30 active AI bills made it through. That number matters more than any single launch this week, because there is still no federal standard sitting above it.

Two days earlier we wrote about the federal review order that got pulled hours before signing. The structural consequence of that decision is playing out right now in Sacramento. With no national framework, the rules that actually bind AI products in the United States get written by whoever moves first, and California is the largest market moving fastest.

The clock is the story. The Legislature adjourns for summer recess on July 2, returns August 3, and runs to a sine die date of August 31. So a bill that cleared crossover still has to survive policy committees, an appropriations suspense file, and a floor vote in the second chamber inside a four-week window, then a second sprint in August, then the governor's pen. Last year that final filter cut hard: Newsom signed SB 53 on September 29 and vetoed several other AI bills the same season. Crossover is the halfway mark, not the finish line.

SB 53 Was the Model Layer. This Crop Is the Deployment Layer.

Here is the framing that makes the 2026 bills legible. SB 53, the Transparency in Frontier Artificial Intelligence Act that Newsom signed last September, regulates the model layer. It applies to large frontier developers with annual revenue above $500 million, which in practice is five to eight companies: OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, and whoever else crosses the line. It is about training compute, governance frameworks, transparency reports, and critical safety incident disclosure, enforced by the California Attorney General at up to $1 million per violation.

The bills that just cleared crossover are aimed somewhere else entirely. They regulate the deployment layer: the chatbot a company points at its customers, the automated system that screens a job applicant, the AI scribe in a therapy session, the disclosure on a real estate listing. SB 53 governs the handful of labs that build the models. This crop governs the tens of thousands of businesses that ship products on top of them. If you operate an agent or a consumer-facing AI feature, the second category is the one that lands on you.

The Bills Worth Tracking

Thirty bills is too many to read closely, so here are the ones with the clearest reach into products people are actually shipping. Status reflects action as of the May 29 crossover gate.

A few patterns jump out of that list. Companion chatbots are the single most legislated category in the building, with at least four separate bills (AB 2023, SB 1119, SB 300, SB 867) coming at the same product from different angles: child safety, explicit content, and a flat ban on putting one inside a toy. That is what a moral panic looks like when it reaches a legislature, and the consumer-assistant cohort that just added action permissions is the target whether the vendors read it that way yet or not.

The second cluster is automated decision systems. SB 947 and SB 719 are the descendants of the employment and ADS bills that stalled in 2025, and they are back. If your agent makes or informs what California calls a consequential decision (hiring, housing, credit, healthcare access), the disclosure and accountability rules are being drafted in this window, not next year.

The third is the quiet one: the natural-person bills. Separate measures specify that a public school employee and a CSU instructor have to be a human being, not an AI system. It reads like housekeeping until you notice the legislature is now writing into statute the places where a model is not allowed to stand in for a person at all. That is a different kind of line than a disclosure rule, and it is worth watching where it spreads.

California Is Not Alone, Which Is the Problem

The same week California cleared crossover, the multistate picture filled in. Illinois gave final approval to SB 315, a frontier model safety act the governor has said he will sign, which makes it the second state after California to legislate the model layer directly. Colorado was expected to sign a chatbot safety bill and a separate measure restricting AI in psychotherapy. Louisiana sent three AI bills to its governor as it wrapped its session. Thirty-some states introduced AI legislation this year.

For an operator, the through-line is not any one state's rule. It is that the federal vacuum guarantees a patchwork, and a patchwork means you comply with the strictest applicable standard across every state you touch, because geofencing a chatbot's safety behavior by user location is both technically awkward and a bad look in a deposition. California sets the practical floor for the same reason its emissions rules set the floor for cars: it is too big to build a separate product for, so its rules become everyone's rules by default. That dynamic was the entire point of the federal order that got killed. Scrapping the one shot at preemption did not produce less regulation. It produced fifty venues writing it at once, with the biggest one moving fastest.

Our Take

The labs spent the spring optimizing for the model layer of compliance, mapping their safety stacks to SB 53 and the EU AI Act. That work is real and it is mostly done. The layer that is still wide open is the one almost no one outside California is tracking week to week: the deployment rules that govern the products built on top of the models. Those are the rules that decide whether your customer-facing agent needs a human-disclosure banner, whether your hiring screen needs an audit trail, whether your healthcare scribe can run at all.

If you ship an AI feature to California users, the practical move this month is unglamorous: pull the three or four bills above that map to your product, read the actual obligations rather than the headlines, and assume the strictest version passes. The ones that clear the second chamber by July 2 are the ones that reach Newsom's desk in the fall. The veto pen is real, but planning around a veto is not a compliance strategy.

Three Signposts Over the Next Ninety Days

The July 2 survivor list. Which of the companion-chatbot and ADS bills actually clear the second chamber before recess. A bill that stalls in appropriations is functionally dead for the year, so the early-July count is the real shortlist of what reaches the governor.

The companion-chatbot consolidation.Whether the four overlapping chatbot bills get merged or amended into one another, or whether the legislature ships conflicting definitions that vendors then have to reconcile. The drafting collisions are where compliance cost actually lives.

The Illinois and Colorado signatures.Whether a second and third state lock in their own model-layer and chatbot rules on a different schedule than California. Two states with slightly different frontier-safety statutes is the moment the patchwork stops being theoretical and starts being a build constraint.

The capital story got the headlines this spring. The regulation story is the one that changes what you are allowed to ship, and it is being written in a four-week sprint that most of the industry is not watching. We will update this as the July 2 deadline approaches.