California lawmakers approve independent testing rules for frontier AI models

Legislation authored by Senator Jerry McNerney directs state officials to establish an auditor qualification system by 2028.

California lawmakers approved legislation to establish a state framework for independent evaluators to test frontier artificial intelligence models for safety risks before deployment.¹² The bill, designated Senate Bill 813 and authored by Senator Jerry McNerney, cleared its final votes on August 30, 2026, sending the measure to Governor Gavin Newsom for consideration.²³

Under the legislation, the California Government Operations Agency must establish application requirements, qualification criteria, and disciplinary procedures for outside auditing entities known as Independent Verification Organizations by January 1, 2028.³⁴ The measure aims to create standard methods for third-party evaluators to inspect model safeguards rather than relying entirely on internal safety claims made by artificial intelligence developers.²³

The policy debate surrounding external evaluation comes as the practical costs of auditing complex models continue to escalate.¹ Analyzing modern frontier systems often requires deploying separate automated models to evaluate system behaviour.¹ In one notable incident, a model evaluation and testing investigation conducted by the safety research group METR into an OpenAI and Hugging Face security issue required the equivalent of $400,000 in artificial intelligence tokens, an expense paid by OpenAI.¹

SB 813 builds on California's previous legislative efforts to structure oversight for advanced computing systems.²³ In September 2025, state lawmakers passed Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, which Newsom signed into law on September 29, 2025.⁵⁴ That statute required large developers with annual revenues exceeding $500 million and models trained using at least 10^26 floating-point operations to publish risk mitigation frameworks and report critical safety incidents to the Office of Emergency Services.⁵⁴

While SB 53 focused on mandatory disclosures and operational incident reporting by developers, SB 813 addresses external evaluation standards.²³ The newly approved bill directs state officials to convene working groups composed of technical experts and engineers from competing technology companies to help draft auditing criteria.³⁴ These criteria must align where practical with existing regulatory and professional assurance standards.³

The auditing structure established under SB 813 remains voluntary for software developers.³ The measure does not mandate that technology companies submit models to an Independent Verification Organization before release, nor does it establish legal liability solely for failing to meet a designated standard.³ If an artificial intelligence system causes harm, a completed audit performed according to the framework may serve as relevant evidence in court, but it does not provide conclusive legal immunity.³

To maintain evaluator independence, the bill prohibits designated verification organizations from having financial, managerial, or operational dependence on the businesses they assess.³ Independent verification entities may accept compensation from assessed companies at normal market rates, but contracts cannot condition payment on the final outcome of an evaluation.³ Designated auditors must also submit annual reports on methodologies and governance starting 12 months after receiving state approval, while retaining unredacted records for five years.³

The legislation was sponsored by Fathom, a policy and technology organization that developed the auditing model.²³ Andrew Freedman, chief executive of Fathom, said the framework allows qualified outside experts to evaluate model risks while remaining flexible enough to keep pace with changing technology rather than locking rigid requirements into statute.²³

Reporting note: this piece draws on reporting from the Semafor policy review published September 4, 2026, legislative analysis from Unite.AI, and state legislative tracking records.

Source: Reed Albergotti via Semafor, September 4, 2026.

References

This article is based on 6 sources, listed in the order they are cited.

1. 1 RA Reed Albergotti announcement · 4 Sep 2026 View: US needs nimble approach to technological governance See the source
2. 2 IS ICT Solutions & Education third party · 2 Sep 2026 California Legislature Passes AI Safety Bill Establishing Independent Verification Framework See the source
3. 3 U unite.ai California Legislature Passes Independent AI Safety Verification Bill See the source
4. 4 AL AI Law Tracker third party · 10 Jul 2026 AI Laws in California (CA) — Compliance Guide 2026 See the source
5. 5 F fisherphillips.com third party · 15 Sep 2025 California Lawmakers Pass Landmark AI Transparency Law for Frontier Models: How SB 53 Differs from Last Year’s Failed Attempt See the source
6. 6 G govbuddy.com California SB 813: Independent verification organizations See the source