California's 2025 frontier AI law makes the biggest developers publish safety plans and report incidents

On September 29, 2025, in Sacramento, Governor Gavin Newsom signed Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, written by Senator Scott Wiener of San Francisco. Companies that train the most powerful AI models must publish how they handle catastrophic risks, report the most serious safety incidents to the state and leave their safety staff free to raise the alarm. More than a year on, the state runs a reporting portal for the law, while the federal government has pressed to limit what states can demand of AI developers.

What came before

SB 53 is California's second attempt. Wiener's SB 1047 of 2024 would have set up a broader regulatory framework for the largest models to prevent catastrophic harms. Newsom vetoed it and, in September 2024, convened a working group of AI experts whose leaders included Stanford's Fei-Fei Li. Its final report, released on June 17, 2025, endorsed no bill but argued for a “trust, but verify” approach. The Assembly's privacy committee described SB 53 as “much narrower” than its predecessor, relying on disclosure rather than prescribing safety standards.

Who the law covers

The duties turn on two thresholds, as the committee's analysis described the bill as amended on September 5, 2025. A frontier developer is one that has trained, or started to train, a foundation model with more than 10²⁶ integer or floating-point operations of computing power. A large frontier developer is one whose annual gross revenue, together with its affiliates, exceeded $500 million in the previous calendar year. The committee noted that amendments made after its earlier vote had raised that revenue line from $100 million and cut the maximum civil penalty from $10 million to $1 million.

The central idea is catastrophic risk: that a frontier model materially contributes to the death or serious injury of more than 50 people, or more than $1 billion in property damage, in a single incident, by giving expert-level help with chemical, biological, radiological or nuclear weapons, acting without meaningful human oversight in a cyberattack or what would be murder, assault, extortion or theft if done by a person, or escaping its developer's or user's control.

What developers must do

Large frontier developers must write, follow and publish on their websites a frontier AI framework. The attorney general's office says it must describe how the company builds national and international standards and industry best practice into its approach, sets and tests thresholds for dangerous capabilities, applies mitigations, responds to critical safety incidents and manages the risks of using its own models internally. Every frontier developer must also publish a transparency report when it releases a new or substantially modified frontier model; large developers add summaries of their risk assessments, according to the committee analysis. Wiener's office said disclosures may be redacted to protect intellectual property.

According to the committee analysis, critical safety incidents go to the Governor's Office of Emergency Services (Cal OES) within 15 days of discovery. If an incident poses an imminent risk of death or serious physical injury, the developer must tell an appropriate authority, such as a law enforcement agency, within 24 hours. Wiener's office pointed to a provision it called a world first: a developer that catches a model lying about how well its safety controls work, in a way that materially increases catastrophic risk, must report that as well. Large developers also send Cal OES confidential summaries of their assessments of catastrophic risk from internal use, every three months or on another reasonable schedule they set.

Employees responsible for managing these risks may report a developer to the attorney general if they have reasonable cause to believe its activities pose a specific and substantial danger to public health or safety from a catastrophic risk, or that it has broken the act, and developers may not gag or retaliate against them. Wiener's office stressed that the law creates no new liability for harm done by AI systems. It also provided for CalCompute, a planned public computing cluster.

Supporters and critics

“California has proven that we can establish regulations to protect our communities while also ensuring that the growing AI industry continues to thrive,” Newsom said in a statement. In its analysis of the September 5 version, the Assembly committee recorded opposition from the Silicon Valley Leadership Group and the Chamber of Progress, while the California Chamber of Commerce, the Computer & Communications Industry Association and TechNet opposed it unless amended; some groups, it noted, might not yet have reacted to the latest changes.

Wiener's office contrasted the law with the EU's AI Act, under which safety plans go privately to a government agency (see our explainer on the EU's 2025 code of practice); SB 53's disclosures are less detailed, it said, but public.

The federal push, and what applies now

An executive order that President Donald Trump signed on December 11, 2025, said state-by-state regulation creates “a patchwork of 50 different regulatory regimes.” It told the attorney general to set up a task force to challenge state AI laws, asked the Commerce Department to identify onerous ones, including laws that may compel disclosures in breach of the First Amendment, and told the Federal Communications Commission to consider a federal disclosure standard that would preempt conflicting state laws. In March 2026, the White House asked Congress to preempt unduly burdensome state AI laws, arguing that states “should not be permitted to regulate AI development.”

California has resisted. On December 18, 2025, Attorney General Rob Bonta joined 23 attorneys general in telling the FCC that it lacks the authority to preempt state AI laws. In a September 23, 2026, letter to congressional leaders, a group of state attorneys general that included Bonta asked for federal safety rules for frontier AI that would not preempt state laws, and cited SB 53 among the state laws under which they say they are holding AI labs accountable. On September 9, 2026, Newsom signed two follow-up bills: SB 813, which sets out a framework for independent organizations that can assess whether AI systems and models comply with state law, and AB 1405, which creates a state registry of AI auditors.

SB 53 remains on the books. The attorney general's office takes employee reports through a dedicated category in its online contact form, and Cal OES's portal accepts incident reports and risk-assessment summaries from developers and the public. According to the committee analysis, Cal OES, the attorney general and the Department of Technology must report to the Legislature on the law each year from January 1, 2027.