The California Report on Frontier AI Policy is a 53-page report published June 17, 2025 by the Joint California Policy Working Group on AI Frontier Models, commissioned by California Governor Gavin Newsom's office to provide an evidence-based framework for frontier AI policymaking. Rather than endorsing or opposing specific legislation, it derives eight policy principles organized around a "trust but verify" ethos: balancing innovation with material risk reduction through transparency and third-party verification rather than prescriptive restrictions.
The report was co-led by Jennifer Tour Chayes (UC Berkeley), Mariano-Florentino Cuéllar (Carnegie Endowment), and Li Fei-Fei (Stanford), with lead writing by Rishi Bommasani (Stanford) and Scott R. Singer (Carnegie Endowment). It was written for and acknowledged by Governor Newsom.
Summary of argument
The report covers the evidence base for risk assessment, transparency as a governance tool, whistleblower and third-party evaluator protections, adverse event reporting systems, and threshold design for policy interventions. Its central ethos, "trust but verify," rejects a "prove safety before deployment" model in favor of "deploy with transparency plus ongoing verification." The report does not advocate for or against specific legislation; it instead frames eight principles that state officials could use to craft AI law.
Eight policy principles
- Balance benefits and material risks. Targeted interventions should weigh transformative potential against severe, potentially irreversible harms. The report describes California as the epicenter of global AI innovation, concentrating both the upside and the responsibility.
- Ground AI policymaking in empirical research. Use observed harms, predictions from technical methods, historical case comparisons, modeling, simulations, and adversarial testing, rather than only reactive incident response.
- Early design choices create path dependencies. The report argues that the foundational governance architecture shapes long-term trajectories, drawing on case studies from internet governance, and that proactive risk assessment before harms materialize is essential.
- Build a robust and transparent evidence environment. Require industry to publish information about systems, informed by clear standards. The report cites case studies from consumer products and energy in which leveraging industry expertise alongside independent verification worked.
- Greater transparency addresses current information deficits. The report states that the AI industry has not coalesced around transparency norms for foundation models, describing "systematic opacity" in key areas, and that disclosure enables informed decision-making for consumers, the public, and future policymakers.
- Whistleblower protections, third-party evaluations, and public-facing information sharing are key transparency instruments. Carefully tailored policies can target current gaps in data acquisition, safety practices, pre-deployment testing, and downstream impacts.
- Adverse event reporting systems enable monitoring of post-deployment impacts. Even well-designed safety policies cannot prevent all adverse outcomes; the report points to existing regulatory authorities (FDA, CPSC, NHTSA models) as precedents for incident reporting.
- Policy intervention thresholds should align with sound governance goals. Compute, revenue, and deployment-scale thresholds should be calibrated to the risk rather than set arbitrarily. The report examines evidence for threshold design and cites SB 53's 10^26 FLOP plus $500M revenue threshold as one calibration.
Key framing choices
The report frames the foundation model as the regulated unit, treating frontier and foundation models as the entry point for governance much as the EU AI Act's GPAI category does, and characterizes downstream deployment regulation as harder to target effectively. It presents industry expertise as a resource rather than an obstacle, citing case studies from aviation safety and pharmaceutical regulation in which industry-informed standards paired with independent verification outperformed purely adversarial regulatory approaches.
The Working Group draws researchers from UC Berkeley, Stanford, and the Carnegie Endowment rather than think tanks close to the industry. Draft reviewers included Yoshua Bengio, Arvind Narayanan, Helen Toner, Lawrence Lessig, and others across the ideological spectrum.
Relationship to California AI legislation
The report was produced in June 2025, after the veto of SB 1047 (September 2024) but before the signing of SB 53 (September 2025). It provided the academic framework for the transparency-first approach that SB 53 embodied. The "trust but verify" framing directly influenced the SB 53 approach (California SB 53 — Transparency in Frontier AI Act). Specific alignments the report's analysis maps onto SB 53 include:
- SB 53's "trust but verify" ethos matches principle 5
- SB 53's 10^26 FLOP threshold matches the report's threshold design discussion
- SB 53's whistleblower protections match principle 6
- SB 53's quarterly internal-use assessment matches the adverse event reporting system in principle 7
The report is not a legislative blueprint for SB 53 and does not endorse any specific bill; it provides evidence-based academic backing for the policy choices SB 53 made.
Provenance and reception
The report was produced by a recognized academic working group at the request of a sitting governor, with wide review by AI policy scholars. The co-lead roster includes Fei-Fei Li (ImageNet, Stanford HAI), Jennifer Tour Chayes (Berkeley), and Mariano-Florentino Cuéllar (Carnegie Endowment, former California Supreme Court Justice). Confidence in the page is rated high on the basis of this provenance and review.
Relationships
- supports: California SB 53 — Transparency in Frontier AI Act — provides academic framework backing SB 53's approach
- supports: Frontier Compliance Framework (February 2026) — transparency-first, third-party verification model
- supports: AI Safety Cases and Frameworks — "trust but verify" framework contribution
- related: Safety Cases for Frontier AI — safety cases are the academic-policy complement from the UK side
- related: California SB 1047 — Safe and Secure Innovation for Frontier AI Models Act (enrolled + veto) — the vetoed predecessor bill this report partly responded to
- related: Fei-Fei Li — Stanford HAI co-director and Working Group co-lead
- related: Jennifer Tour Chayes — UC Berkeley I-School Dean; Working Group co-lead
- related: Mariano-Florentino Cuéllar — Carnegie Endowment President / former CA Supreme Court Justice; Working Group co-lead
- related: Rishi Bommasani — Stanford CRFM; lead writer
- related: Stanford CRFM — academic center-of-gravity for the report
- related: Carnegie Endowment for International Peace — Cuéllar's current home; Scott Singer affiliation
- related: Yoshua Bengio — draft reviewer
- related: Arvind Narayanan — draft reviewer (principal author of AI as Normal Technology)
- related: Lawrence Lessig — draft reviewer
- related: Helen Toner — draft reviewer