California Gov. Gavin Newsom signs into law two bills, backed by Anthropic and OpenAI, regulating how outside groups evaluate AI for safety
Context & Ripple Effects
California’s approach has moved from Newsom’s veto of SB 1047’s model-focused framework toward requirements tied to disclosed safety practices: SB 53 established AI safety-testing disclosures in 2025, and a 2026 executive order attached safety and privacy guardrails to state AI contracts. The two new laws extend that arc to the outside groups that assess model safety.
Anthropic and OpenAI backing the measures gives the framework support from two companies subject to the broader safety-policy debate, while the governor has framed the laws as strengthening transparency, accountability, and independent assessments.
First-order effects
- Outside organizations conducting AI safety evaluations in California face new legal rules for how those assessments are performed, while Anthropic, OpenAI, and other evaluated developers gain a defined state framework for using them.
- California adds independent evaluations and audits to the accountability mechanisms surrounding the safety-testing disclosures required under its 2025 AI safety law.
Second-order effects
- AI developers selling into California or seeking to demonstrate responsible safety practices have an incentive to align their assessment processes with the state’s rules, making third-party evaluation more central to compliance planning.
- OpenAI’s public call for capability-based federal safety requirements puts California’s state-level approach in sharper contrast with a national standard that leading labs say they want.
Third-order effects
- If other jurisdictions treat California’s framework as a reference point, independent AI evaluation can become a regulated assurance function rather than a voluntary lab-led practice.
- The policy direction favors AI labs able to document testing, engage outside assessors, and meet government procurement guardrails—an emerging model of state-mediated AI governance.
The trend: AI governance is shifting from broad arguments over frontier-model rules toward operational oversight of testing, disclosure, and independent assessment.