California’s Senate Bill 813 is a novel approach to AI safety

California is taking steps to address artificial intelligence (AI) safety with Senate Bill 813, a proposal that brings together industry, civil society, and third-party evaluators to develop clear standards. The bill aims to encourage companies to adopt AI-safety practices by offering limited liability protections to developers who can demonstrate a genuine, documented effort to comply with standards set by independent evaluators.

By encouraging voluntary cooperation between government and industry, the bill creates a predictable framework that manages AI risks without stifling innovation or relying on heavy-handed regulation, preserving the flexibility and incentives of a free market. However, lawmakers must be cautious to avoid prematurely creating expectations about what regulations should exist and who certifies that companies are genuinely making a good-faith effort.

Senate Bill 813, proposed by state Sen. Jerry McNerney (D–Pleasanton), outlines a framework for AI safety certification. It authorizes the attorney general to designate financially independent private organizations—such as nonprofits, industry consortia, or accredited standards bodies— as multistakeholder regulatory organizations (MROs) for three-year terms to certify AI models and applications against risks such as cybersecurity and nuclear threats. Certified developers receive an affirmative defense against civil lawsuits for AI-related harm. MROs must be financially independent and submit annual public reports on their evaluations and risk mitigation efforts.

Rather than relying on litigation to shape AI accountability, the bill attempts to build a voluntary process that rewards cooperation and early action. SB 813 focuses on a preventative model based on risk assessments and certified mitigation steps. This structure aims to create predictability for developers and greater public assurance, without overburdening innovation with frivolous lawsuits.

Artificial intelligence is advancing rapidly, and many of the most pressing concerns remain unresolved. Early regulation often moves faster than our understanding of the technology. For instance, as Reason Foundation has noted, California’s Assembly Bill 1064 proposes restrictions on chatbots that use “emotional language,” despite a lack of clarity about what that means or how such behavior should be measured. The result is a regulation that overreaches without clear standards (the bill is still under consideration by lawmakers).

Using independent evaluators is not without risks, however. SB 813 draws on models from other industries, such as the Financial Industry Regulatory Authority (FINRA), which oversees companies and individuals who buy and sell stocks and bonds. FINRA was created in 2007 but traces back to earlier voluntary associations formed under federal law. Over time, the federal government delegated formal rulemaking and enforcement powers to these organizations, supervised by the Securities and Exchange Commission (SEC). While still technically private, FINRA now plays a central regulatory role in U.S. financial markets.

In 2024, the U.S. Court of Appeals for the D.C. Circuit held that FINRA had exceeded the authority it had when it expelled a brokerage without adequate government oversight. The court found that allowing a private entity to wield enforcement powers normally reserved for public agencies can violate separation-of-powers principles. This decision underscores the risk that, if California’s MROs gain enforcement powers without clear legislative guardrails, they could face similar challenges.

MROs could serve as a useful bridge between the government and the expanding network of AI companies. As the field grows, many firms are developing internal safety measures, such as red teaming (simulated attacks to find vulnerabilities) or adversarial testing (feeding malicious inputs to test robustness), without clear standards or coordination with public agencies. MROs can help fill this gap by developing shared benchmarks for transparency and offering a way to assess whether companies are taking emerging threats seriously.

The bill’s approach may serve as a viable alternative, but lawmakers should proceed carefully so it does not inadvertently give coercive power to organizations that are meant to be voluntary.

The post California’s Senate Bill 813 is a novel approach to AI safety appeared first on Reason Foundation.