On September 9, 2026, California Governor Gavin Newsom signed two bills that regulate not artificial intelligence itself, but the people and firms who inspect it. Senate Bill 813, authored by Senator Jerry McNerney, and Assembly Bill 1405, authored by Assemblymember Rebecca Bauer-Kahan, together create what the state describes as the first framework in the country for certifying who is legally allowed to audit AI systems. The move shifts California's approach one layer up the stack: rather than only writing rules for AI, the state is now building the profession that will check whether those rules are followed.
OpenAI
Anthropic
Google
MicrosoftWhat each bill actually does
The two laws do different jobs that fit together. SB 813 sets up a category of "independent verification organizations," bodies that can assess AI systems and models for compliance with state law. It directs California's Government Operations Agency to publish the criteria those organizations must meet by January 1, 2028, including disclosure of their qualifications, methods and testing tools, and a demonstration that they can manage conflicts of interest. Crucially, SB 813 is structured as a voluntary track rather than a blanket mandate: the statute does not require every AI company to hire a state-designated verifier.
AB 1405 is the licensing half. It stands up an online AI Auditor Registry and, from January 1, 2029, generally bars any unregistered person or organization from conducting an AI audit that is required to assess compliance with state law. Registered auditors have to deliver reports that spell out the audit's scope, objectives, findings and supporting documentation, along with any deficiencies they found, recommended fixes, and the limits of what they were able to test.
The independence problem the laws are trying to solve
The most technically interesting provision is the one on payment. Under SB 813, an auditor may be paid by the company it is assessing, but its compensation cannot depend on the findings, and it must keep operational and managerial independence from that company. Anyone who has watched other assurance markets will recognize the design. It is the same structural tension that financial auditing spent a century trying to resolve: the entity being examined is also the client writing the check, and the quality of the examination erodes whenever a bad grade threatens the relationship.
California is not regulating AI here so much as regulating the mirror the industry will be asked to look into.
On what SB 813 and AB 1405 actually govern
That framing matters because it sets a realistic bar for what these laws can deliver. An audit regime is only as credible as the independence, methods and enforcement behind it. The bills gesture at all three, but the substance lands in rulemaking that has not happened yet. Until the Government Operations Agency publishes its criteria, the market cannot know how demanding registration will be, how conflicts will be policed, or what an audit must actually test to count.
A staged rollout, not a switch
The dates are the part most worth internalizing. Nothing binding happens on day one. The verification criteria are due at the start of 2028, and the registry gate does not close until the start of 2029. That gives developers, prospective auditors and the state itself more than two years to stand up an ecosystem that does not currently exist at scale.
A two-year runway before the rules bite
The laws are signed, but the machinery that gives them force arrives in stages through the end of the decade.
Newsom signs SB 813 and AB 1405, the first US framework to certify who is allowed to audit AI systems.
SB 813 directs the Government Operations Agency to publish criteria for independent verification organizations.
AB 1405 stands up the AI Auditor Registry. From this date an unregistered party generally cannot conduct a covered AI audit.
Dates as specified in the enacted bills and state guidance. Later rulemaking will fill in operational detail.
Reading the gap between signing and effect is the analytically honest way to size a law like this. A common mistake in AI-policy coverage is to treat the signature as the moment behavior changes. In practice, the signature starts a clock, and the interesting decisions, which methodologies qualify, how independence is verified, and what happens to a company that fails, are made during the runway. Whether these laws bite in 2029 will depend far more on the rulemaking than on the text signed in September.
Scope: broader than the frontier labs
It would be easy to read this as a frontier-model story, but the reach is wider. Coverage extends to organizations deploying AI in consequential settings such as hiring, insurance and other services where an automated decision affects a person's access to opportunity. That is a deliberate choice. The systems most likely to produce discrimination or safety complaints are frequently not the largest chatbots but the quieter models embedded in screening, pricing and eligibility. By defining the auditor rather than only the audited, California is building a mechanism that can in principle be pointed at any of them.
The federalism subtext is unmistakable. In signing, Newsom repeated a line he has used before, urging Washington to act and positioning California as filling a vacuum. That posture is also a risk. A single large state writing assurance rules for a national industry invites both a patchwork of state regimes and a future federal preemption fight, and the compliance burden of that uncertainty tends to fall hardest on smaller deployers.
What it means for teams building on AI
For most companies the near-term takeaway is not panic but posture. An audit regime rewards systems that can be inspected: clear records of which model produced which output, what data and prompts were involved, and how a decision could be reconstructed after the fact. Organizations that treat a single proprietary model as an opaque dependency will find that harder than those that keep their AI layer portable and observable, able to route across providers and to export the evidence an auditor will ask for. That is the same discipline that protects against vendor lock-in and price shocks, and it is the philosophy behind model-agnostic platforms like Metir, where the point is to keep the assistant, its records and its outputs under the user's control rather than trapped inside one vendor's stack. Auditability, in other words, is less a compliance chore you bolt on in 2029 than a property you design in now.
California has not answered whether third-party AI auditing will work. It has built the container the answer will eventually go in. The next two years of rulemaking will decide whether that container holds something with teeth or becomes another registry that exists mostly on paper. Either way, the state has drawn the line where the hardest questions in AI accountability actually live: not in the model, but in who gets to grade it.
Sources:
- Governor Newsom signs first-in-the-nation AI safeguards (Governor of California, Sept 9, 2026)
- California Starts Regulating the People Who Audit AI (PYMNTS, Sept 2026)
- Newsom signs legislation establishing framework for third-party AI auditors (StateScoop)
- Newsom Signs AB 1405 Creating California's First AI Auditor Registry (Startup Fortune)
- California signs first US AI audit law (Tech Times, Sept 10, 2026)
- California Gov. Newsom signs two bills to create nation's first AI auditing framework (Transparency Coalition)
Image credits
- Hero: California State Capitol, Sacramento, June 2019. Photo by Frank Schulenburg, Wikimedia Commons, licensed CC BY-SA 4.0.
- In-body: Official portrait of Gavin Newsom, State of California, Wikimedia Commons, public domain.