California waters down AI disaster bill before final vote

California lawmakers just blinked in their high-stakes standoff with Silicon Valley.

Faced with fierce pushback from tech giants, venture capitalists, and members of Congress, state Senator Scott Wiener pushed through several key amendments to SB 1047, the state’s landmark AI safety bill. The changes heavily reflect advice from AI frontier lab Anthropic and easily passed the California Appropriations Committee on Thursday ahead of a final vote.

SB 1047 still sets out to prevent catastrophic AI disasters—specifically scenarios that kill large numbers of people or cause over $500 million in cybersecurity damages by holding frontier developers liable. But the updated version drastically dials back the state’s enforcement teeth.

Softer penalties and a pass for open-source tuners

The biggest shift takes away preemptive litigation powers from the state. Under the original bill, the California attorney general could sue AI labs for negligent safety practices before any disaster occurred. Following Anthropic's feedback, the AG can now only seek injunctive relief to halt risky operations beforehand, saving actual lawsuits for after a catastrophe happens.

Lawmakers also scrapped plans for an entirely new bureaucracy called the Frontier Model Division. Instead, they folded a 9-person Board of Frontier Models into an existing state agency to issue safety guidance and set compute thresholds.

Criminal liability is off the table as well. AI companies previously had to certify their safety evaluations under penalty of perjury; now, they only need to submit public statements detailing their safety practices. The required standard of care dropped from providing "reasonable assurance" to showing "reasonable care."

Crucially for the broader ecosystem, the amendments create a safe harbor for open-source tweaking: anyone spending under $10 million fine-tuning a covered model is explicitly shielded from developer liability under the bill, leaving the legal burden on the original creator.

Silicon Valley is still pushing for a veto

The watered-down version was clearly designed to placate critics and give Governor Gavin Newsom something less controversial to sign. Yet many in the industry remain unimpressed.

Andreessen Horowitz general partner Martin Casado dismissed the changes on X as "window dressing" that fails to address the bill's real flaws. Moments after the amendments passed committee, eight California members of the U.S. House of Representatives sent a joint letter urging Newsom to veto the legislation, arguing it harms startups, science, and the state's economy.

Even Anthropic hasn't formally endorsed the bill, stating it is still reviewing the final text since not all of its proposals made the cut. SB 1047 now heads to the California Assembly floor, then back to the Senate, before finally hitting Newsom’s desk.

Why it matters

Wiener made a smart tactical retreat here, but the core ideological war over AI regulation is far from resolved.

By carving out small open-source fine-tuners and ditching the perjury threats, the amendments effectively disarm the argument that this bill instantly kills indie AI development and puts open-source hackers in handcuffs. For developers spending modest compute to customize open weights, the $10 million threshold provides immediate breathing room.

However, the central friction point remains untouched: holding frontier model makers legally liable for extreme catastrophic outcomes. As long as that core liability exists, major tech labs will view California as a hostile testing ground for frontier intelligence. We think Newsom now faces an impossible balancing act between protecting California’s crowning tech boom and setting the standard for global AI safety.