California's Governor Newsom vetoes SB 1047 and the most consequential US AI bill of the year dies short of his desk
A bill that would have imposed safety obligations on the largest training runs and indemnified whistleblowers passed both chambers comfortably. The veto letter took the trouble of saying why.
On 29 September 2024, California Governor Gavin Newsom vetoed Senate Bill 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act. The bill had passed the State Senate by 32 to 1 and the Assembly by 41 to 9. It would have imposed a duty of reasonable care on developers of frontier AI models, defined as models trained with more than ten to the twenty-sixth FLOPs and costing more than one hundred million dollars to train. The bill would also have created legal protections for AI-lab whistleblowers and a Frontier Model Division within the state's Government Operations Agency.
The veto letter was unusual for its specificity. As reported by the Los Angeles Times and Politico the same evening, Newsom rejected the bill not on procedural grounds but on substantive ones: the regulation by compute threshold, he wrote, was insufficiently sensitive to the actual deployment context of an AI system, and risked imposing identical obligations on a research model and a public-facing customer-service system. The letter explicitly noted that he supported AI regulation in general and signed seventeen other AI-adjacent bills that same week.
What the bill would have done
The substantive critique was that the bill regulated by training-compute threshold rather than by use case. Open-source advocates including a16z's Andreessen Horowitz, Meta and Hugging Face had argued through the summer that the threshold would chill open-weight releases by exposing developers to liability for downstream use. Anthropic, in a notably nuanced position taken by Dario Amodei in a public letter to Newsom in August, had taken a softer line, saying the bill was net-positive but in need of amendments around the kill-switch and liability provisions.
Vetoed for what it would not catch, not for what it would.
The veto closed off the most aggressive state-level frontier-AI bill of the 2024 cycle. It also reset the political calculation for similar legislation in other states. By 2026, state-level frontier-AI bills tend to regulate by deployment context rather than training compute, an approach that draws explicitly on Newsom's veto letter.
Discussion