California AI Regulation

California AI regulation is the body of state law governing the development and deployment of artificial intelligence in California, the home state of most United States frontier-model developers. Since 2024 it has produced the country's most consequential AI statutes, including a frontier-model transparency law, SB 53, and disclosure duties for training data.^[1]

History

California began legislating on AI in 2024 with a package of bills. Two of them set the trajectory for everything that followed.

AB 2013, signed on 28 September 2024 and effective from 1 January 2025, requires developers of generative AI systems to publish documentation about their datasets — sources, whether personal information was used, and how the data was collected and cleaned.^[2]

SB 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, cleared both chambers of the legislature but was vetoed by Governor Gavin Newsom on 29 September 2024. His veto message argued that regulating only the most expensive, largest-scale models "establishes a regulatory framework that could give the public a false sense of security" while leaving smaller but potentially dangerous models unregulated.^[3] The Governor subsequently convened a working group on frontier-model policy to reshape the approach.

The result was SB 53, the Transparency in Frontier Artificial Intelligence Act (TFAIA), authored by Senator Scott Wiener and signed on 29 September 2025 — a deliberately pared-down successor to SB 1047, focused on transparency and incident reporting rather than pre-licensing.^[4]

Key Provisions

SB 53 applies to a large developer: an organisation that trained at least one foundation model using more than 10^26 floating-point operations of compute and earned over US$100 million in annual revenue. The definition holds until 1 January 2027, after which the Attorney General may replace it by regulation.^[5]

The law is anchored to catastrophic risk — a foreseeable and material risk that a developer's work materially contributes to the death or serious injury of more than 50 people, or to more than US$1 billion in damage from a single incident involving a dangerous capability. Dangerous capabilities are specified in the statute: expert-level assistance with chemical, biological, radiological or nuclear weapons, conducting or assisting cyberattacks, assisting serious crimes with limited human intervention, and evading the developer's own controls.^[5]

Three duties follow. Large developers must publish a transparency framework describing how national, international and industry-consensus standards are built into their practices; publish a safety and security protocol explaining how they will protect model weights and manage risk; and report critical safety incidents — including unauthorised access to or exfiltration of model weights, harm from a materialised catastrophic risk, loss of control of a model, and deceptive circumvention of controls. Enforcement sits with the California Attorney General through civil penalties.^[4][5]

Why It Matters for AI

SB 53 is the first United States state law to regulate frontier models directly, and its threshold gives other jurisdictions a template far narrower than the European Union's risk tiers. It also marks a shift in US policy style: rather than approving models before release, it requires developers to publish how they govern risk and disclose when something goes wrong. Whether other states copy it, or federal pre-emption overrides it, remains open.^[6]

>Key Takeaways

• California sets the pace for US AI law; SB 53, signed in September 2025, is the first state frontier-model statute.
• Duties are transparency-based — publish a framework, publish a safety protocol, report critical safety incidents.
• Scope is narrow: over 10^26 FLOP of training compute and US$100 million in annual revenue.
• Prior vetoes and later federal pre-emption questions mean the rules are still moving.

See Also

• EU AI Act
• Colorado AI Act
• US Executive Orders on AI
• AI Regulation in the United States
• Malaysia AI Governance Framework
• AI Chip Export Controls