CaliforniaSB 532025-2026 Regular SessionSenateWALLET

Artificial intelligence: frontier models.

Your PRIA Score

Score Hidden

Personalized for You

How does this bill affect your finances?

Sign up for a PRIA Policy Scan to see your personalized alignment score for this bill and every other piece of legislation we track. We analyze your financial profile against policy provisions to show you exactly what matters to your wallet.

Free to start

Bill Overview

Analyzed Economic Effects

6 provisions identified: 3 benefits, 1 costs, 2 mixed.

Stronger whistleblower rights at AI companies

The law protects covered employees at frontier developers who report serious AI dangers or legal violations. Employers cannot block or punish protected disclosures to the Attorney General, a federal authority, a supervisor, or an internal investigator. Companies must offer an anonymous reporting process, give monthly status updates, and brief officers and directors each quarter. Courts can shift the burden to employers in retaliation suits and can issue fast temporary orders. Employers must give clear notice of these rights each year. If you win a case, the court can award reasonable attorney fees. These remedies add to, not replace, other state whistleblower protections.

Fines and truth rules for developers

Large frontier developers that break these rules face civil penalties up to $1,000,000 per violation. Only the Attorney General can bring a civil action to recover these penalties. Frontier developers also cannot make materially false or misleading statements about catastrophic risk. Large frontier developers cannot misstate their compliance with their safety framework. A good‑faith exception applies to reasonable statements.

Big AI firms must publish safety plans

Large frontier developers must write, follow, and clearly publish a frontier AI safety framework. They must review it each year and publish any material change with a short reason within 30 days. Before or at deployment, all frontier developers must post a transparency report with contact info, release date, languages, outputs, intended uses, and use limits. Large developers must also add summaries of catastrophic‑risk assessments, their results, and the role of third‑party evaluators. Large developers must send OES summaries of internal-use catastrophic‑risk assessments every three months or on another written schedule. Developers may redact published materials to protect trade secrets and security, but must explain the redaction and keep the unredacted records for five years.

Report serious AI incidents to state

The Office of Emergency Services (OES) runs a system for developers and the public to report critical safety incidents. Reports must include the date, why it is critical, a plain description, and whether it involved internal model use. Developers must report within 15 days of discovery, or within 24 hours if there is an imminent risk of death or serious injury. OES reviews reports and can share them, or covered employee reports, with the Legislature, Governor, or other agencies after weighing trade secrets and security risks. These reports are exempt from public records requests. Beginning January 1, 2027, the Attorney General also publishes an anonymized, aggregated annual summary of covered employee reports. OES also accepts confidential catastrophic‑risk assessment summaries from large developers and limits access to staff who need to know.

State plans public cloud for research

The state creates a consortium to plan a public cloud cluster called CalCompute to support safe, ethical, and equitable AI research. It aims to locate the platform at the University of California, with state‑owned hosting and staff to operate and train users. A 14‑member group, serving unpaid but reimbursed for expenses, develops the plan. The report is due by January 1, 2027, and the consortium dissolves after it submits the report. If CalCompute is at UC, UC may accept private donations to run it. This section only operates if the Legislature funds it.

Defines frontier AI, firms, and risks

A frontier model is one trained with more than 10^26 operations, including fine‑tuning. A large frontier developer is a frontier developer and its affiliates with over $500 million in revenue last year. A catastrophic risk is a single incident likely to cause over 50 deaths or more than $1 billion in damage, with listed exclusions. Critical safety incidents include certain deadly harms, loss of control, and deceptive techniques outside testing. Beginning January 1, 2027, the Department of Technology reviews these definitions each year. The law also states that loss in equity value is not property damage under this chapter.

Sponsors & Cosponsors

Sponsor

Scott Wiener
Democratic • Senate

Cosponsors

Susan Rubio
Democratic • Senate

Roll Call Votes

All Roll Calls

Yes: 208 • No: 17

Senate vote • 9/13/2025

Item 153 — Senate SFLOOR

Yes: 29 • No: 8

House vote • 9/12/2025

Item 1000 — Assembly AFLOOR

Yes: 59 • No: 7

legislature vote • 9/11/2025

Vote in CX32

Yes: 12 • No: 1

legislature vote • 8/29/2025

Vote in CX25

Yes: 11 • No: 1

legislature vote • 7/16/2025

Vote in CX32

Yes: 10 • No: 0

legislature vote • 7/1/2025

Vote in CX13

Yes: 12 • No: 0

Senate vote • 5/28/2025

Item 226 — Senate SFLOOR

Yes: 37 • No: 0

legislature vote • 5/23/2025

Vote in CS61

Yes: 6 • No: 0

legislature vote • 4/21/2025

Vote in CS61

Yes: 6 • No: 0

legislature vote • 4/8/2025

Vote in CS53

Yes: 13 • No: 0

legislature vote • 3/25/2025

Vote in CS48