The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
California’s SB 1047 was never a law. The Legislature passed the proposed Safe and Secure Innovation for Frontier Artificial Intelligence Models Act in 2024, but Governor Gavin Newsom vetoed it on September 29. The bill would have imposed safety and security obligations on developers of certain very large, powerful AI models; it would not have penalized a company every time a chatbot produced a wrong or offensive answer. California later enacted a different frontier-AI measure, SB 53, in 2025.
What SB 1047 proposed
SB 1047, sponsored by California State Senator Scott Wiener during the 2023–2024 legislative session, was designed for so-called frontier AI models: systems developed with exceptionally large amounts of computing power and money. It was not a general rule for every software company or every interaction with a chatbot. The California Legislature’s record confirms the bill’s final status: vetoed by the governor.
Coverage often summarized the proposal with a development-cost threshold of more than $100 million, alongside computing-power criteria. That shorthand does not capture the full statutory definition or its changes during the legislative process. The enrolled bill text is the right source for the final proposed language; the practical point is that the bill targeted a narrow class of highly resourced model development, not AI in general.
The proposal would have required covered developers to establish and follow written safety and security protocols, assess and test risks, and provide for emergency shutdown or disabling of a model. It also contemplated independent audits in a future implementation period and protections for employees who disclosed safety concerns. The often-used phrase “kill switch” refers to one emergency-control element, not the whole bill.
Its enforcement provisions contemplated civil actions and penalties for specified violations and serious harms or imminent public-safety risks. The bill named harms such as death or bodily injury, property harm, theft or misappropriation, and threats to public safety. It did not create a simple rule that any harmful output automatically made the model developer liable.
#1 Best Overall
Who would have been responsible?
SB 1047 mattered because it tried to place some safety responsibility upstream, on developers of the most powerful models, rather than leaving accountability solely with the person who later used a system. It also addressed computing providers in limited circumstances. That is different from saying that every company whose product uses AI would have been responsible for every downstream result.
- User liability: A person who deliberately misuses a model may be responsible for that conduct.
- Developer liability: The model creator may have control over training, safeguards, evaluations, and known risks.
- Deployer liability: A business integrating a model may control the setting in which it is used, such as a hiring, health-care, or financial workflow.
- Shared responsibility: In practice, control of a risk can be distributed among the developer, deployer, provider, and user.
The bill’s approach was most significant for developers of covered models, but the policy debate raised harder boundary cases: a cloud provider may furnish computing without controlling a model’s use; an open-source model may be altered and redistributed; and a general-purpose model can be safe in one setting but risky when embedded in a high-stakes system.
Why AI companies objected
Major AI companies, including OpenAI, opposed the bill. As contemporary reporting described, OpenAI argued that the proposal threatened California’s AI economy and could encourage engineers and entrepreneurs to leave the state. Other objections centered on several connected concerns:
- Uncertain liability: Critics worried that companies could be penalized for harms caused by downstream users, malicious actors, or model behavior that is difficult to predict. The final text did not amount to automatic strict liability for every hallucination, but critics disputed whether its standards and exposure were sufficiently clear.
- Open-source effects: Opponents said obligations primarily suited to large laboratories could burden smaller developers and people who release or modify models openly. That was a criticism of possible effects, not proof that every open-source project would have been covered or forced to close.
- Innovation and competition: Companies warned that compliance costs and legal uncertainty could slow development or shift investment elsewhere. Those were predictions, not established outcomes.
- State-by-state rules: Industry advocates favored a consistent federal framework and warned that different state requirements could fragment the market. Supporters countered that states should not have to wait for Congress to address serious risks.
These arguments were not simply “companies versus safety.” They concerned which systems should be covered, how to measure risk, where liability should fall, and whether a state-level framework could be both effective and workable.
Why supporters wanted it
Supporters argued that the most capable models could potentially assist with large-scale cyberattacks, biological threats, fraud, or other serious harms. Developers have access to the systems and resources needed to test for misuse and strengthen safeguards, they said, so they should have enforceable duties rather than relying only on voluntary promises. Documentation, evaluation, independent scrutiny, and protected reporting could make it harder to ignore risks as products are developed and released.
The argument was about severe and potentially large-scale risks—not a claim that every chatbot error is catastrophic. A model can generate misinformation, offensive content, or unsafe instructions without causing the kinds of harms contemplated by the bill. Conversely, a model that behaves acceptably in ordinary use could still present risks when adapted, integrated into critical systems, or deliberately exploited.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #3
What “accountable when AI does bad stuff” meant—and did not mean
The original headline captured the political conflict but compressed the legal question. SB 1047 would have established obligations for developers of certain covered models and allowed civil enforcement for specified violations and serious harms or imminent threats. It was not a blanket penalty for an inaccurate answer, rude response, or any output someone considered harmful.
Nor does the distinction mean harmful outputs never matter. An output could be part of a chain of events leading to harm, and regulators or courts would have to apply the relevant legal standards to the facts. But a user’s misuse, a product’s design, a deployer’s decisions, and a developer’s safeguards are different questions. The bill sought to make some of those responsibilities explicit for frontier-model development; it did not resolve every question of causation or responsibility across the AI supply chain.
There are practical limits to any safety requirement. A shutdown mechanism may disable a centrally operated service, but it cannot necessarily retrieve copies already downloaded or stop a separately deployed modification. A written protocol can document a process without guaranteeing meaningful tests. Those edge cases help explain why both the bill’s potential benefits and its implementation burdens were contested.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why Newsom vetoed the bill
Newsom said the proposal was well-intentioned but not the best approach. In his official veto message, he objected that the bill focused on the size of a model rather than adequately distinguishing whether an AI system was used in a high-risk environment, made critical decisions, or handled sensitive data. In other words, he questioned whether model scale alone was the right basis for the proposed framework.
That objection reflects a genuine regulatory trade-off. A size-based approach can identify powerful systems before every use is known, but size is an imperfect proxy for real-world risk. A use-based approach can focus on consequential settings, but risks may emerge before deployment or move across contexts. Newsom’s veto did not mean California would avoid AI regulation altogether: his office announced other safe-and-responsible-AI initiatives at the time.
What happened after the veto
California returned to frontier-AI legislation through a different measure. Newsom signed SB 53, the Transparency in Frontier Artificial Intelligence Act, on September 29, 2025. It is a later law with a different title and framework; it is not SB 1047 revived or enacted under another name.
The accurate timeline is therefore simple: SB 1047 passed the California Legislature in 2024, was vetoed on September 29, 2024, and never became law. California later enacted SB 53 in 2025. The headline’s political fight was real, but calling SB 1047 a “new law” was wrong even at the time it was published, when the proposal was still pending.
The underlying policy dispute
The lasting disagreement was not whether AI systems should have any safeguards. It was about how to define the most consequential risks, whether the developer or the deployer should bear responsibility, how to treat open and shared models, and whether state rules can protect the public without creating unworkable uncertainty. SB 1047 was one attempt to answer those questions for the largest frontier models. Its veto ended that bill, not California’s debate over how to govern AI.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
