FPF and Brookings — California SB 53 Compliance Analyses (Oct–Dec 2025)

Two practitioner-facing analyses of California SB 53 (Transparency in Frontier Artificial Intelligence Act, TFAIA), summarized together because they describe the same law from adjacent angles. The Future of Privacy Forum (FPF) analysis, by policy analyst Justine Gluck, focuses on the legal-compliance mechanics. The Brookings Institution analysis, by Malihe Alikhani and Aidan T. Kane, focuses on implementation uncertainty and the political economy of enforcement.

Scope and covered entities

Both sources concur on the law's definitions. A frontier developer is one that trains a foundation model using more than 10^26 integer or floating-point operations, including compute for subsequent fine-tuning, reinforcement learning, or material modifications. A large frontier developer is a frontier developer whose annual gross revenue, with affiliates, exceeds $500M in the preceding calendar year. Brookings estimates the current scope at roughly 5–8 major companies, naming OpenAI, Anthropic, Google DeepMind, Meta, and Microsoft.

The law defines catastrophic risk as a foreseeable material risk that the model could cause 50 or more deaths or $1B or more in damages; provide expert-level assistance with weapons of mass destruction; engage in autonomous criminal conduct; or evade developer control.

Compliance obligations

FPF organizes the law's requirements into four obligations:

1. Frontier AI Framework (large developers) — annual publication of catastrophic-risk identification and mitigation, governance, cybersecurity, and standards alignment. Redactions are allowed for trade secrets and national security.
2. Transparency reports (all frontier developers) — published before deployment, covering intended uses, modalities, restrictions, risk assessments, and third-party evaluator documentation.
3. Critical safety incident disclosure — to the California Office of Emergency Services, on a 15-day standard timeline or a 24-hour timeline for imminent events. Covers unauthorized tampering, materialized catastrophic risk, loss of control, and deliberate safeguard evasion.
4. Whistleblower protections — no retaliation, required employee notice, and anonymous reporting channels.

Brookings adds that the law's quarterly regulatory summaries create continuous oversight obligations on top of the annual-framework cycle.

For practitioners, FPF recommends building written governance frameworks covering risk, mitigation, and cybersecurity aligned to national and international standards; pre-deployment reporting templates covering capabilities, uses, restrictions, and third-party evaluation summaries; incident-classification systems that distinguish 24-hour critical events from 15-day standard events; anonymous reporting channels paired with written whistleblower policies; multi-jurisdictional planning, especially around NY RAISE Act divergences; and tracking of California Department of Technology threshold recommendations.

Enforcement

Enforcement authority rests with the California Attorney General, with penalties of up to $1M per violation, scaled by severity. FPF flags as a distinctive design choice that the law grants no rulemaking power: the Department of Technology may only recommend updates to thresholds, subject to Legislative approval, in contrast to SB 1047's AG rulemaking model.

Evolution from SB 1047

FPF tabulates how SB 53 differs from the vetoed SB 1047:

The kill-switch entry stands in tension with the characterization on legislation/california-sb-53 that the law retains kill-switch requirements. FPF asserts the kill-switch mandate was removed, whereas the statutory source page describes "reasonable safeguards including the ability to promptly shut down frontier models." The divergence may reflect that the original SB 1047 kill-switch was stronger (mandatory, pre-training, with explicit testability requirements) and the SB 53 language is substantially weaker. Both characterizations can hold if "kill switch" is read as a spectrum; the point is flagged for future reconciliation rather than resolved here.

Comparison with the NY RAISE Act

FPF contrasts SB 53 with the NY RAISE Act across several dimensions: RAISE applies an "unreasonable risk" liability standard against SB 53's catastrophic-risk language; RAISE scopes by $100M+ compute while SB 53 scopes by $500M+ revenue; RAISE penalties run $10M–$30M against SB 53's $1M cap; SB 53 requires public incident reporting while RAISE uses 72-hour windows; and SB 53 includes employee protections that RAISE lacks.

Brookings on compliance burden and open questions

Brookings advances the interpretive thesis that "much of the law codifies practices that major companies already claim to follow," implying that compliance cost may be lower than the public narrative suggests for existing frontier labs, while still creating real cost for any new entrant crossing the threshold. The analysis is described in the wiki as engaging the "business reality gap" theme, since codifying existing practice implies both low marginal burden and limited marginal safety gain.

Brookings raises four open practitioner questions: whether transparency reports will carry meaningful information or drift toward boilerplate; whether regulators can keep pace with monthly model-evolution cycles; whether whistleblower channels will surface actionable insights; and whether federal action will harmonize with, or override, state experiments.

The analysis also offers three critiques. The first is a regulatory-cliff concern: the thresholds create a structural incentive to stay just below 10^26 FLOP or $500M revenue, a threshold-gaming risk. The second is federal-preemption pressure: OpenAI and Google publicly push federal preemption to avoid state fragmentation, a dynamic connected to techno-federalism and broader federal-versus-state tensions. The third is an enforcement-gap concern: Brookings characterizes the $1M cap as "lighter" than the EU AI Act or the NY RAISE Act's $30M.

Relationships to other governance instruments

The two analyses add practitioner compliance detail to legislation/california-sb-53 and its source summary. They connect to techno-federalism through Governor Newsom's framing of SB 53 as "a blueprint for other states," and provide a concrete divergence table against the NY RAISE Act. Read alongside the AI LEAD Act and the Colorado AI Act, they support a compliance-practitioner framing for comparing US regulatory approaches. The Brookings "codifies existing practice" thesis also relates to Amodei's framing of SB 53 as graduated regulation.

Confidence is high on scope, compliance obligations, enforcement mechanics, and the SB 1047 deltas, where both sources concur and match the statutory text. It is lower on the "kill switch removed" versus "kill switch retained" question, flagged above for reconciliation, and on Brookings' "codifies existing practice" thesis, which is a contested interpretive claim rather than a factual observation.

Relationships

• supports: California SB 53 — Transparency in Frontier AI Act (legislation page)
• supports: California SB 53 — Transparency in Frontier AI Act (source summary)
• related: New York RAISE Act (S. 8828) — direct comparison
• related: California SB 1047 — Safe and Secure Innovation for Frontier AI Models Act (enrolled + veto) — predecessor evolution
• related: Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race — state-led governance thesis
• related: The Adolescence of Technology — Amodei's framing of SB 53 as graduated regulation

Sources

• Future of Privacy Forum — Gluck, "California's SB 53: The First Frontier AI Law, Explained" (Oct 3, 2025; updated Feb 4, 2026).
• Brookings — Alikhani & Kane, "What is California's AI safety law?" (Dec 23, 2025).