California Vetoes SB 1047: Newsom Halts the AI Safety Bill That Split Silicon Valley
On September 29, 2024, Governor Newsom vetoed SB 1047 — America's most aggressive frontier AI safety bill — pausing a fight that had torn Silicon Valley apart.
On September 29, 2024, California Governor Gavin Newsom vetoed SB 1047 — then America's most aggressive frontier AI safety bill, which would have required large-model developers to run safety testing and bear liability for catastrophic harms.
The bill had split Silicon Valley: Hinton, Bengio and Musk backed it; OpenAI, Google, Meta and much of the VC world opposed it. Newsom's veto message argued it targeted big models rather than high-risk uses, while he signed milder bills like AB 2013 on training-data transparency.
SB 1047's defeat set the tone for U.S. AI regulation: a federal vacuum, a state-by-state patchwork, innovation first. It also seeded California's 2025 second attempt via the transparency-focused SB 53.
The Real Divide: Regulate Models or Regulate Uses
The fight was never over whether to regulate but where to aim. The bill drew its line at training cost ($100M), holding developers of above-threshold models liable for catastrophic harms — supporters argued only choking the source contains tail risk; opponents warned it would turn open-weight releases into self-inflicted legal exposure, hitting startups and academia first. Newsom's veto message bought the second narrative: regulate by the risk of actual deployment scenarios, not model scale. That stance unexpectedly echoes the EU's use-based risk tiers (see our EU AI Act coverage) — while leaving the question of whether frontier models themselves need a threshold suspended in mid-air.
The coalition map was the striking part: of deep learning's founding trio, Hinton and Bengio backed the bill while the commercial labs opposed it almost unanimously — the safety-versus-acceleration rift went public here and has resurfaced in every regulatory clash since.
Our Take
In hindsight, the veto was the moment US AI regulation chose its lane: no federal statute, state-level trial and error, industry self-governance by default — a posture the White House AI Action Plan later pushed to its limit (see our coverage). The lasting lesson: absent regulation does not mean absent risk. Copyright litigation and minor-protection cases are becoming the de facto substitute regulators (see our NYT lawsuit and Character.AI coverage) — where legislatures stay silent, courts and public opinion fill the gap.
This article aggregates official announcements and public reporting; original sources are linked below.
Source:加州州长办公室否决信