California SB 1047
California SB 1047 was the proposed Safe and Secure Innovation for Frontier Artificial Intelligence Models Act. Its final enrolled version would have imposed specified safety-and-security protocol, audit, incident-reporting, whistleblower, and shutdown-capability duties on developers of covered high-compute models and would have created frontier-model governance bodies and CalCompute provisions. The Legislature passed the bill in 2024, but the Governor vetoed it, so SB 1047 did not create operative legal duties.
Origin and context
SB 1047 was introduced on 7 February 2024 and amended repeatedly before the Legislature enrolled the final text on 3 September. Public debate often focused on shorthand such as a model kill switch, developer liability, and a $100 million training-cost threshold, but the text contained qualifications, compute criteria, control boundaries, and multiple institutional provisions. Brookings analyzed several contested descriptions against the amended bill. On 29 September, Governor Newsom vetoed it, arguing that its threshold-centered design could miss risky smaller models and apply stringent standards to covered models even when used in lower-risk contexts.
Why it matters
Although it did not take effect, SB 1047 became a major reference point in debates over whether frontier-model regulation should attach duties to developers before deployment, how coverage should use compute and cost thresholds, and what role audits, safety protocols, shutdown capabilities, and civil enforcement should play. Its legislative path also helps explain later California policy. For researchers and governance teams, the bill is best treated as a dated design proposal whose exact version matters, not as shorthand for all California AI safety policy or as proof of a current legal requirement.
Example
A policy comparison might ask how the enrolled SB 1047 would have classified a large training run and which safety protocol or audit duties would have followed. The analyst should cite the final enrolled version, state that the bill was vetoed, and separate its proposed duties from those later enacted in SB 53. A compliance memo should not instruct a company to satisfy SB 1047 as current California law; it may instead use the bill to trace policy alternatives and legislative history.
How it differs
California SB 53 / TFAIA
SB 53/TFAIA is a different bill enacted in 2025. It shares frontier-AI governance themes but uses a different structure and cannot be described as the final version of SB 1047. SB 1047 remains the vetoed 2024 proposal.
Maturity and evidence
Maturity is rated 3 because SB 1047 reached a stable, enrolled legislative text and generated substantial independent analysis, but it was vetoed and never became law. The inherited maturity of 5 is therefore inappropriate: legal codification did not occur. Its historical influence is real, while its practical provisions remain those of a failed bill rather than a regulated lifecycle category.
Limits and open questions
This entry is not legal advice and does not summarize every version of the bill. Claims made during the 2024 debate may refer to earlier amendments, while the page's substantive description uses the final enrolled text. Cost figures alone do not reproduce the statutory covered-model definition. The Governor's veto message explains the executive decision but is not a neutral empirical evaluation of every provision; Brookings offers independent analysis but also advances an interpretive position. Current California duties must be checked in laws that were actually enacted.
Related terms
References
- SB-1047 Safe and Secure Innovation for Frontier Artificial Intelligence Models Act.California Legislative Information · 2024-09-03 · class A
- Bill History: SB-1047 Safe and Secure Innovation for Frontier Artificial Intelligence Models Act.California Legislative Information · 2024-02-07 · class A
- SB 1047 veto messageOffice of Governor Gavin Newsom · 2024-09-29 · class A
- Misrepresentations of California's AI safety billBrookings Institution · 2024-09-27 · class B
Last updated: 2026-09-05
This term is also covered in the Skills Atlas as ai risk management skill.
This term is also covered in the Skills Atlas as model evaluation skill.