Anthropic says revised California AI bill likely does more good than harm

Silicon Valley's biggest AI labs have spent months railing against California's landmark safety bill. Now, one of the loudest players has broken ranks.

Anthropic CEO Dario Amodei sent a letter to California Governor Gavin Newsom stating that the revised SB 1047 bill is substantially improved, to the point where its benefits likely outweigh its costs. He added a caveat, though: the company is not completely certain, and several provisions still feel concerning or ambiguous.

A lone voice among the frontier giants

The proposed legislation, spearheaded by Democratic State Senator Scott Wiener, targets frontier AI models that cost over $100 million to train or require a defined threshold of computing power. It requires frontier developers to conduct safety testing, maintain an effective kill switch to shut models down if they go awry, and gives the state attorney general the authority to sue non-compliant companies.

Most big tech players have pushed back hard. Meta wrote to Wiener warning that the bill threatens to make California an inhospitable place for AI development. Google also voiced serious reservations. OpenAI sent its own letter, arguing that AI should be regulated at the federal level and warning that California's piecemeal effort creates legal uncertainty that could slow progress.

Anthropic took a different route. After previously submitting a Support if Amended proposal, the company saw several of its recommendations adopted by the Assembly Appropriations Committee. Most notably, the updated draft eliminated a contentious plan for a dedicated government AI oversight committee.

Focusing on catastrophic risks

In his letter, Amodei pointed out that Anthropic's work with biodefense and cybersecurity researchers suggests that catastrophic misuse could become a reality within the next 1 to 3 years. Because the updated threshold targets only models requiring massive development budgets, smaller startups and developers are largely exempt.

Amodei highlighted three main wins in the bill. First, it compels frontier labs to implement transparent Safety and Security Protocols. While Anthropic, Google, and OpenAI have voluntary frameworks, some competitors have been vague or made questionable claims about their safety audits.

Second, the bill clarifies tort law by tying a foundation model's downstream liability directly to the rigor of its safety protocol. Third, it financially incentivizes labs to invest in the nascent science of risk reduction, rather than pouring every dollar into raw model capability. Still, Amodei warned against government overreach, urging regulators to maintain a laser focus on genuine catastrophic risks.

Why it matters

This split marks a significant turning point in the AI policy debate. For everyday users and independent developers, the high cost threshold means current open-source tooling and standard fine-tuning workflows won't be disrupted by the red tape.

Strategically, Anthropic's move looks very calculated. Having branded itself around safety from day one and invested heavily in safety frameworks, Anthropic is turning its existing compliance overhead into a universal industry requirement. It forces rivals like Meta and OpenAI to adopt similar, legally binding guardrails.

The real risk, however, is a fragmented regulatory landscape. As OpenAI pointed out, state-by-state rules can create massive legal gray areas. If California's rules prove too burdensome, frontier labs might simply route compute clusters elsewhere, testing whether safety regulation can actually keep up with an borderless race.