Newsom Vetoes California’s SB 1047 AI Regulation Bill

On September 29, 2024, Governor Gavin Newsom vetoed SB 1047. In his official veto message, he accepted the need for safeguards but argued that focusing them on large, expensive models could omit smaller systems deployed in dangerous contexts; that is an argument about scope, not evidence that risk is absent.

The decision matters beyond California. The state is home to many of the companies, investors, and labs leading the race to build generative AI models. A state law could have become a practical benchmark for the industry, even without being a federal regulation.

A law focused on frontier models

SB 1047, sponsored by state Senator Scott Wiener, was formally known as the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act. Its aim was to establish requirements for developers whose models required more than 10^26 floating-point or integer operations to train. The bill also covered models whose development cost more than $100 million. Source

The proposal required those companies to develop safety protocols, test whether their models could enable critical harm, and retain the ability to shut down the system if necessary. Such harms included the creation of chemical, biological, radiological, or nuclear weapons, cyberattacks with serious consequences, and other large-scale catastrophe scenarios.

The bill was therefore not intended to regulate every application that uses AI. A customer-service chatbot, image generator, or document-analysis tool would not automatically fall within its scope. The legislation targeted the most expensive foundation models—systems that can later be adapted and distributed for multiple uses.

It also called for the creation of a specialized public division focused on frontier models and mechanisms allowing California’s attorney general to act against violations. Supporters argued that once these systems reach particularly dangerous capabilities, voluntary codes of conduct will not be enough.

The problem with measuring risk by cost

Newsom has agreed with the goal of strengthening safety but rejected the mechanism chosen. In his veto message, he argued that focusing on the largest and most expensive models could create a false sense of security.

The governor said a relatively small model could be used in a high-risk setting—for example, to make critical decisions or process sensitive data—while a much larger model could have less dangerous applications. Training size, money invested, and computing power are relevant indicators, but they do not by themselves explain what a system can do or how it will be deployed.

That is one of the central debates in AI regulation. Some rules seek to control the technology based on its capabilities, while others focus on how it is used: a model that summarizes text is not the same as one integrated into an energy infrastructure, healthcare network, or defense system. The European Union’s AI Act, for example, combines obligations tied to high-risk uses with specific requirements for certain general-purpose models.

Support and opposition in Silicon Valley

SB 1047 had divided the industry. AI safety organizations and figures such as Geoffrey Hinton and Yoshua Bengio supported the bill, arguing that it established proportionate safeguards for a technology with growing capabilities.

Several technology companies, investors, and industry groups, however, warned that the law could raise barriers to entry for smaller companies, penalize the development of open models, and push business activity out of California. Anthropic backed a modified version of the bill, while OpenAI and Meta raised objections to the proposal.

The disagreement was not purely commercial. It also reflected a difficult question: how to require preparedness for risks that remain relatively uncommon without turning that preparedness into a burden only tech giants can afford.

California will seek an alternative framework

The veto does not mean California is abandoning artificial intelligence regulation. Newsom has tasked experts, including researcher Fei-Fei Li, with developing an empirical framework for analyzing the capabilities and risks of frontier AI systems. The state is also working on assessments of the safe use of AI within its own government agencies.

The governor has also signed other initiatives related to transparency and protection against artificially generated content. The difference is that SB 1047 sought to intervene directly in the development of the most advanced models before they reached the market.

The debate will return. The veto nullifies this version of the bill, but it does not resolve who should oversee models capable of performing increasingly complex tasks, according to what criteria, or before what kind of deployment. California is thus keeping open a debate that Washington and Brussels are watching closely.

The veto message and bill answer different questions

The legislative text shows duties and definitions; the veto explains why the governor declines to make them law. Evaluation requires placing them side by side. Criticism of a threshold does not erase protocols, audits or powers in other sections, and supporting the purpose does not mean accepting the mechanism.

The proposed alternative framework deserves the same discipline: owner, timetable, resulting document and authority. Commissioning a report starts a process; it does not replace an enforceable duty. When recommendations appear, distinguish draft, regulation, voluntary guidance and law.

Size and use can be combined

A regime can impose documentation duties on frontier developers and additional controls on critical deployments. The first layer uses administrable indicators; the second addresses context and consequence. Treating them as mutually exclusive weakens regulatory design.

Ask who can verify. A self-assessment without method or incident access asks for trust; full publication may expose sensitive information. Protected audit, aggregate reporting and authority to demand records are intermediate solutions that need to be written into the text.

The transferable skill is to follow policy from proposal to effect: version, passage, veto, alternative, regulation and evidence. A veto closes one legislative vehicle, not the public question. The timeline prevents intention being presented as a rule or no law as no work.

A veto does not settle the whole policy

The message explains why the governor rejected that text and which alternative he supported at that time. It does not turn his arguments into technical findings or prevent similar duties returning elsewhere. Each mechanism is the unit of analysis, not a coalition’s temporary victory.

Track which elements survive: evaluations, incident notices, whistleblower protection, supervisory authority or thresholds. Then look for them in later legislation, voluntary agreements or agency rules. A policy may disappear as a bill number and persist in pieces.

Independent evaluation needs a precise commission

“External audit” does not guarantee independence or quality. Readers need to know who pays, what access the evaluator receives, whether the laboratory can exclude results, which methods are published and how vulnerabilities are handled. Without those conditions, the label describes a commercial relationship, not verification.

The final criterion is traceability: proposal, version, decision, alternative and later evidence. That sequence tests whether criticism produced a better design or merely moved the decision, without attributing intentions the file does not document.

When another proposal appears, the table reveals what genuinely changed and what merely received a new name. That continuity is more useful than retelling the story from scratch in every political cycle.