California SB 53 — Transparency in Frontier Artificial Intelligence Act
Summary
Section titled “Summary”Senate Bill 53 — California Transparency in Frontier Artificial Intelligence Act (TFAIA), drafted by State Senator Scott Wiener (D-SF) and signed by Governor Newsom on 2025-09-29, enters into force on 2026-01-01.
Significance: the first US state law specifically targeting safety and transparency for frontier AI models. Following SB-1047 (vetoed by Newsom in 2024), California enters the frontier-AI regulatory space in a more measured but more focused form.
Definition of “Frontier Model”
Section titled “Definition of “Frontier Model””“A foundation model trained on a quantity of computing power greater than 10²⁶ integer or floating-point operations (FLOPs)”.
→ The same threshold used in Biden EO 14110 (now rescinded). Covers:
- the GPT-4o / o3 / o4 series (OpenAI)
- the Claude Opus / Sonnet series (Anthropic)
- the Gemini Ultra series (Google DeepMind)
- the Llama 3+ series (Meta)
- the Grok series (xAI)
“Large Frontier Developer”
Section titled ““Large Frontier Developer””Frontier-model developers with annual revenue ≥ $500M bear heightened obligations.
Core obligations
Section titled “Core obligations”1. Frontier AI Framework disclosure
Section titled “1. Frontier AI Framework disclosure”Large frontier developers must publish an AI Framework on the company website describing:
- How it incorporates national standards, international standards, and industry-consensus best practices.
- Specific references: NIST AI RMF, ISO/IEC 42001, MLCommons, etc.
- Alignment with the White House Voluntary Commitments, the EU GPAI Code of Practice, the Frontier Model Forum, and similar frameworks.
2. Transparency reports on model releases
Section titled “2. Transparency reports on model releases”On every new model release or material modification:
- Release date.
- The model’s modalities (text / vision / audio / multimodal).
- Intended uses.
- Any deployment limitations.
- Summary and results of the catastrophic-risk assessment (large developers only).
3. Critical safety-incident reporting
Section titled “3. Critical safety-incident reporting”- Report potential critical safety incidents to the California Office of Emergency Services (OES).
- Covering: biological weapons, cyber-attacks, loss of model-autonomy control, large-scale manipulation, etc.
4. Whistleblower protection
Section titled “4. Whistleblower protection”- Employees and contractors who disclose significant health or safety risks of frontier models receive civil and criminal protections.
- Employers are prohibited from retaliating against whistleblowers.
Penalties
Section titled “Penalties”- Up to $1,000,000 per violation (for large developers with annual revenue ≥ $500M).
- The California Attorney General has sole enforcement authority.
- No private right of action.
Comparison with SB-1047 (same author — Wiener)
Section titled “Comparison with SB-1047 (same author — Wiener)”| Dimension | SB-1047 (2024, vetoed) | SB 53 (2026, in force) |
|---|---|---|
| Core | Mandatory safety obligations + “kill switch” + reporting | Transparency + framework disclosure |
| Scope | ”Covered model” (compute + $100M training cost) | “Frontier model” (compute alone) |
| Type of obligation | Substantive capability controls | Documentation + disclosure |
| Penalties | High (including state AG injunctive relief) | Up to $1M / violation |
| Political fate | Vetoed by Governor Newsom 2024-09 | Signed by Governor Newsom 2025-09 |
Strategy: a retreat from “mandatory obligations” to “transparency”, avoiding the reason given for Newsom’s veto (“stifles California innovation”) while preserving the regulatory value of “government access to systematic information”.
Corporate response
Section titled “Corporate response”Anthropic (headquartered in California)
Section titled “Anthropic (headquartered in California)”Published its SB 53 Compliance Framework (2025-10):
- Publishes Usage Policy + RSP + Model Card as the Framework.
- Internalizes the CA OES critical-safety-incident reporting process.
- Updates whistleblower policy.
OpenAI / Google DeepMind
Section titled “OpenAI / Google DeepMind”- Compliance plans are expected to be merged with the EU GPAI Code of Practice (which likewise covers transparency + safety).
- A “single documentation set” strategy spanning EU / CA is gradually taking shape.
Non-US companies
Section titled “Non-US companies”- Mistral, DeepSeek and others are equally subject if their models cross the 10²⁶ FLOP threshold + are made available in California.
Conflict with EO 14365
Section titled “Conflict with EO 14365”SB 53 is one of the primary targets of EO 14365 (2025-12):
- EO 14365 tasks the AI Litigation Task Force with challenging “unduly burdensome” state laws.
- The California Attorney General has explicitly signalled continued enforcement.
- Federal lawsuits are expected in 2026-Q1/Q2.
Nevertheless: EO 14365 cannot independently preempt state law (it requires Congressional legislation or a Supreme Court ruling); SB 53 continues to take effect in the near term.
Primary text and archives
Section titled “Primary text and archives”| Source | Link |
|---|---|
| Full text (California legislature) | leginfo.legislature.ca.gov/…/SB53 |
| Governor’s signing statement | gov.ca.gov/2025/09/29/…/sb-53 |
| Future of Privacy Forum commentary | fpf.org/blog/californias-sb-53-the-first-frontier-ai-law-explained |
| Brookings commentary | brookings.edu/articles/what-is-californias-ai-safety-law |
| Anthropic compliance framework | anthropic.com/news/compliance-framework-SB53 |
Revision history
Section titled “Revision history”| Date | Event |
|---|---|
| 2024-09 | SB-1047 vetoed by Governor Newsom |
| 2025-02 | SB 53 re-drafted by Senator Wiener |
| 2025-09-29 | Signed by Governor Newsom |
| 2026-01-01 | In force |
| 2026-Q1+ | Federal litigation expected (against the backdrop of EO 14365) |