California pushes new bill to force AI safety disclosures

California lawmakers are taking another swing at regulating artificial intelligence, but this time they are swapping the sledgehammer for a clipboard.

State Senator Scott Wiener just introduced major amendments to SB 53, a revised AI bill that would require the world’s top developers to publish their safety and security protocols and report major safety incidents. If signed into law, California would become the first state to mandate formal transparency from front-runners like OpenAI, Google, Anthropic, and xAI.

Stepping back from SB 1047

Wiener’s previous legislative push, SB 1047, sparked open warfare across Silicon Valley before Governor Gavin Newsom ultimately vetoed it. Newsom subsequently tapped a working group of AI researchers and industry veterans, including Stanford professor and World Labs co-founder Fei-Fei Li, to draft recommendations for state safety policies.

SB 53 leans heavily into that group's findings, which called for a transparent evidence base. Unlike its controversial predecessor, SB 53 completely drops developer liability for model-caused downstream harms. It also explicitly protects smaller players, exempting startups and researchers fine-tuning frontier models or leveraging open-source systems.

The bill additionally introduces legal protections for lab whistleblowers flagging critical risks, specifically defined as incidents involving more than 100 casualties or over $1 billion in damages. To throw the ecosystem a bone, it also proposes CalCompute, a public compute cluster meant to assist resource-strapped academic teams and early-stage builders.

The voluntary honor system is fraying

State pressure is arriving right as voluntary disclosures from leading labs appear to be slipping. Google took months to release a safety paper for Gemini 2.5 Pro after deployment. OpenAI skipped publishing a safety report entirely for GPT-4.1, which independent researchers later found to be less aligned than prior systems.

Big Tech remains divided. Anthropic has cautiously welcomed stronger transparency rules, whereas Meta, Google, and OpenAI continue to push back against state mandates. Meanwhile, broader efforts in Washington to block state-level AI rules for a decade crashed and burned in a 99-1 Senate vote earlier this month, giving momentum to state initiatives like California's SB 53 and New York’s pending RAISE Act.

Former Y Combinator president Geoff Ralston backed the measure, noting that in the absence of serious federal guardrails, California's tailored framework serves as a pragmatic starting point.

Why it matters

For builders and end users, this bill is largely painless. Wiener made sure SB 53 avoids burdening downstream application developers, open-source hackers, or fine-tuners, meaning everyday software workflows will not be tied up in red tape.

We think this pivot is a masterclass in legislative pragmatism. By dropping developer liability and honing in on basic operational disclosure, Wiener effectively disarms the industry argument that regulation will crush California innovation. Forcing labs to publish real protocols and protecting internal whistleblowers targets genuine operational negligence without meddling in code.

The lingering question is enforcement. Disclosure requirements can easily degrade into sanitized PR theater if oversight bodies do not possess the technical competence to audit them. Even so, the era of frontier labs grading their own homework behind closed doors looks increasingly numbered.