California has enacted two laws that set qualification and conduct rules for independent AI auditors, creating public oversight infrastructure without requiring every AI developer to submit every system to an audit. Governor Gavin Newsom signed Senate Bill 813 and Assembly Bill 1405 on September 9, 2026. Together they create a state process for designating expert verification organizations and a registry for people and firms that perform AI audits tied to California-law compliance.
Evidence note: researched September 11, 2026, from the chaptered bill texts, the governor’s announcement, OpenAI’s policy statement, and reporting from Bloomberg Law and Reuters. Toolsfine did not conduct an AI audit or independently test any AI system. Dates and obligations below reflect the enacted text; implementation details can change through agency rulemaking.
At a glance
| Question | Confirmed answer |
|---|---|
| What became law? | SB 813 and AB 1405, approved and chaptered on September 9, 2026. |
| What does SB 813 create? | A state designation process for qualified independent verification organizations by January 1, 2028. |
| What does AB 1405 create? | A public AI Auditor Registry and conduct rules for covered AI audits beginning January 1, 2029. |
| Are all AI companies now required to obtain audits? | No. SB 813 expressly says it does not impose a general audit requirement. |
| Who implements the laws? | California’s Government Operations Agency. |
What California changed
The governor’s September 9 announcement presents the laws as complementary parts of an independent-evaluation framework. The chaptered texts make the division clearer: SB 813 defines how the state may recognize specialized verification organizations, while AB 1405 governs registration and professional conduct for auditors performing a narrower class of compliance audits.
SB 813 requires the Government Operations Agency to establish application requirements, qualification criteria, suspension procedures, and public information for “independent verification organizations,” or IVOs, by January 1, 2028. Applicants must disclose their expertise, proposed benchmarks, metrics, methodologies, and supporting documentation.
The criteria must consider technical competence, credible risk-assessment methods, cybersecurity, recordkeeping, and independence. An IVO may be paid by the organization it evaluates at a reasonable market rate, but payment cannot depend on the result. California must also convene stakeholder working groups and publish the resulting requirements.
What the auditor registry covers
AB 1405 requires a public AI Auditor Registry by January 1, 2029. From that date, an unregistered person or firm cannot offer, sell, or conduct a “covered AI audit.” The law defines that term as an assessment of internal controls, processes, or systems that are necessary for compliance with California law—not every benchmark, red-team exercise, or voluntary product review.
Registered auditors must identify the laws under which they work, list relevant credentials, describe their services, and provide a standard operating procedure that names the standards they use and the basis for reliability claims. Audit reports must state their scope, results, deficiencies, recommended measures where appropriate, evidence gaps, and limitations. Auditors must retain supporting material for at least 10 years and follow independence, objectivity, competence, and conflict-of-interest rules.
The important limitation: audits are not broadly mandatory
SB 813 explicitly says it does not require an AI developer, deployer, or operator to hire an IVO or undergo a covered audit simply to operate in California. It also says that following a state-identified audit standard does not automatically establish or eliminate liability. An audit may be relevant evidence in a harm case, but it is not conclusive.
That distinction matters because “California regulates AI auditors” and “California requires every AI model to be audited” are very different claims. Bloomberg Law likewise characterized the SB 813 designation program as voluntary. Separate laws or contracts may still require a specific organization to obtain an audit; these two measures mainly define who can credibly perform that work and how.
Why the framework matters
AI audits are difficult to compare when evaluators use different definitions, access levels, test sets, and conflict rules. California’s approach targets that assurance layer. Public registration can make credentials and methods easier to inspect, while required disclosures can reveal whether an evaluator had enough access to support its conclusions.
The laws also recognize a structural tension: auditors are often paid by the companies they assess. The new rules do not prohibit that business model, but they require independence controls and bar result-contingent payment. That resembles established assurance practices without treating an AI evaluation as identical to a financial audit.
OpenAI endorsed both bills in a September 9 policy statement, while continuing to call for national rules. Reuters independently reported that endorsement and the company’s broader push for mandatory federal AI-safety requirements. Support from one major developer is relevant industry context, not proof that implementation will be effective or universally accepted.
Practical takeaways
- AI developers: keep traceable safety claims, evaluation artifacts, access records, and conflict disclosures that an outside assessor could examine.
- AI buyers: ask whether a report is a covered compliance audit, a technical evaluation, or a vendor-funded review; the labels are not interchangeable.
- Audit firms: map services to AB 1405’s definition, monitor the registry process, and prepare to disclose methods, credentials, limitations, and independence safeguards.
- Risk teams: do not wait for 2028 or 2029 to define audit scope, evidence access, remediation ownership, and retention periods in contracts.
- Readers: treat state registration as a baseline credential, not a guarantee that an AI system is safe or legally compliant.
What remains uncertain
The statutes leave substantial work to the Government Operations Agency. Final application forms, fees, qualification thresholds, technical standards, and enforcement procedures are not yet available. The agency must consult auditors, developers, startups, civil-society groups, government bodies, and standards organizations, so the practical burden will depend on that implementation process.
The framework also cannot solve every audit problem. Model behavior changes after updates, evaluators may receive incomplete access, and a test can miss rare failures. Buyers and regulators will still need to examine scope, evidence quality, versioning, and known blind spots rather than relying on a registration number alone.
Bottom line
California has built rules for the AI-assurance profession before imposing a universal AI-audit mandate. The immediate effect is institutional, not operational: agencies now have deadlines to define qualified evaluators and a public registry. For developers and buyers, the signal is clear—independent AI claims will increasingly be judged by who performed the work, what evidence they saw, how conflicts were managed, and what the report could not establish.
Sources
- Governor of California: Newsom signs AI safeguards — September 9, 2026.
- California Legislature: SB 813 chaptered text — September 9, 2026.
- California Legislature: AB 1405 chaptered text — September 9, 2026.
- Bloomberg Law: Newsom signs AI-auditor bills — September 10, 2026.
- Reuters: OpenAI pushes for national AI-safety requirements — September 9, 2026.
- OpenAI: The AI policy window is open — September 9, 2026.