AIAcademy · AIAcademy · 2026-05-16
Read SB 53 on leginfo.legislature.ca.gov
The story before the story: SB 1047, Senator Scott Wiener's much-debated 2024 frontier-AI bill, was vetoed by Governor Newsom in September 2024 with an explicit reservation about scope. California then pivoted — not abandoned — and one year later passed something quietly more important.
The Transparency in Frontier AI Act (SB 53) was signed by Newsom on 29 September 2025 and took effect 1 January 2026. It is the first US frontier-specific law on the books, and it works through transparency obligations rather than capability prohibitions. Read in plain English, here is what it actually requires.
Who is covered. A developer is a "large frontier developer" if it trains a model with ≥10²⁶ FLOPs and has annual revenue above $500 million. That two-prong test, by design, captures roughly 5–8 firms — OpenAI, Anthropic, Google, Meta, Microsoft, xAI, and a small number of others — and excludes academic labs and startups.
What they have to publish. Each covered firm must annually publish a Frontier AI Framework documenting governance, risk management, evaluation methodology, and red-teaming. At each major model release, they must publish a Transparency Report describing capabilities, evaluations performed, and known risks. The statute explicitly references the NIST AI Risk Management Framework and ISO/IEC 42001 as the standards anchor — meaning compliance can be measured against published rubrics, not improvised case-by-case.
Incident reporting. Critical safety incidents must be reported to CalOES within 15 days of discovery; imminent threats within 24 hours. Whistleblower protections cover employees who disclose internally or externally. Civil penalties run up to $1 million per violation, enforced by the Attorney General.
The CalCompute companion. A separate provision creates a public-compute consortium intended to give academic and civil-society researchers GPU access at non-market rates. It is the bill's quiet pro-research counterweight.