United States - California - AI Transparency Act (SB 53)

Overview

California Senate Bill 53 (Chapter 138, Statutes of 2025), the Transparency in Frontier Artificial Intelligence Act (TFAIA), represents the most ambitious state-level effort to govern advanced artificial intelligence development in the United States. Authored by Senator Scott Wiener and signed by Governor Gavin Newsom on September 29, 2025, the legislation responds to growing concerns about the potential for frontier AI systems—the most capable and resource-intensive models—to pose catastrophic risks to public safety. The law creates a comprehensive governance framework requiring large developers of frontier models to publish safety frameworks, report risks to government agencies, implement incident response procedures, and protect employees who raise safety concerns. SB 53 takes a targeted approach, applying its requirements only to frontier models (trained with more than 10^26 computing operations) developed by organizations with $500 million or more in annual revenue, concentrating regulatory attention on the entities most capable of producing potentially dangerous AI systems while avoiding burdens on smaller developers, academic researchers, and open-source projects. The legislation establishes California as a global leader in AI safety governance, creating precedents that may influence federal legislation and international frameworks addressing advanced AI risks.

Definitions

SB 53 establishes precise technical definitions that determine regulatory scope. Frontier model means a foundation model trained using more than 10^26 computing operations, including the original training run plus any fine-tuning or modifications. This threshold captures only the largest, most capable AI systems currently in development. Frontier developer means any person who trained or initiated training of a frontier model meeting the computing threshold. Large frontier developer means a frontier developer with annual gross revenues exceeding $500 million from itself and its affiliates combined in the preceding calendar year—the entities subject to most requirements. Catastrophic risk means a foreseeable, material risk of causing death or serious injury to 50 or more people, or property damage of $1 billion or more, resulting from a single incident involving: (1) expert-level assistance in creating, releasing, or using chemical, biological, radiological, or nuclear weapons; (2) conduct without meaningful human oversight constituting cyberattacks or crimes such as murder, assault, extortion, or theft; or (3) the model evading developer control. Importantly, catastrophic risk excludes harms caused by publicly available information, lawful federal government activities, or situations where the AI model did not materially contribute to the harm. Critical safety incident includes unauthorized access to model weights causing injury, materialization of catastrophic risks, loss of model control, or deceptive model behavior that subverts developer controls. Covered employee means an employee whose responsibilities include assessing or managing critical safety risks, the category protected by whistleblower provisions.

Governance and Institutional Framework

SB 53 establishes a distributed governance model leveraging existing California agencies rather than creating new regulatory bodies. The California Attorney General serves as the exclusive enforcement authority, empowered to bring civil actions against violators and impose penalties. The California Office of Emergency Services (Cal OES) receives quarterly internal risk assessment summaries and critical safety incident reports, positioning the state's emergency management agency to monitor AI-related threats alongside natural disasters and other emergencies. The Department of Technology must annually assess the definitions of 'frontier model,' 'large frontier developer,' and 'catastrophic risk' to ensure they remain appropriate as AI technology evolves, providing a mechanism for regulatory adaptation without requiring new legislation. Large frontier developers bear primary compliance responsibility, including publishing frameworks, submitting reports, maintaining whistleblower systems, and implementing their own governance structures. The law explicitly preempts local regulations adopted after January 1, 2025, that specifically regulate frontier developers' management of catastrophic risks, ensuring statewide uniformity. Federal agencies retain concurrent jurisdiction, and SB 53 exempts from disclosure any information that would compromise national security when performed at federal government direction. The governance structure emphasizes transparency through mandatory disclosures rather than pre-approval requirements, allowing continued AI development while creating accountability mechanisms for safety practices.

Key Focus Areas

• Frontier AI Framework Publication: Large developers must publish comprehensive frameworks describing their approaches to identifying, assessing, and mitigating catastrophic risks, incorporating national and international standards and industry best practices.
• Capability Threshold Definition: Frameworks must define and assess thresholds for determining when AI capabilities create catastrophic risk potential, ensuring systematic evaluation of emerging capabilities.
• Risk Mitigation Strategies: Developers must document mitigation measures applied based on risk assessments, creating accountability for safety interventions.
• Third-Party Risk Assessment: Frameworks must address the use of independent evaluators to assess catastrophic risks, introducing external perspectives into safety processes.
• Cybersecurity for Model Weights: Requirements address protection of unreleased model weights—the core trained parameters—from unauthorized access that could enable malicious use.
• Internal Governance Structures: Developers must establish governance ensuring framework implementation, embedding safety responsibilities within organizational structures.
• Quarterly Risk Reporting: Summaries of internal catastrophic risk assessments must be transmitted to Cal OES every three months, enabling government monitoring of identified risks.
• Transparency Reports Before Deployment: Large developers must publish reports detailing model capabilities, safety measures, and intended uses before or concurrent with deploying new frontier models.
• Critical Safety Incident Reporting: Incidents must be reported to Cal OES within 15 days of discovery, or within 24 hours if imminent death or injury risk exists, enabling rapid government response.
• Whistleblower Protection: Robust protections prohibit retaliation against employees reporting safety concerns to authorities, management, or other covered employees with investigative authority.
• CalCompute Public Computing: The law directs development of a public cloud computing consortium to democratize access to AI research resources, addressing compute concentration concerns.

Implementation Framework

SB 53 takes effect upon the Governor's signature on September 29, 2025, with requirements phasing in across multiple timelines. Large frontier developers must immediately establish frontier AI frameworks and publish them on their websites, along with mechanisms for receiving communications from the public and government agencies. Quarterly reporting of internal risk assessments to Cal OES begins immediately, with specific schedules to be coordinated with the Office. Transparency reports must accompany any new frontier model deployment, requiring prospective compliance for future releases. Critical safety incident reporting obligations apply to incidents discovered after the effective date, with 15-day standard reporting and 24-hour expedited reporting for imminent threats. Whistleblower protection provisions require immediate compliance, including posting workplace notices of employee rights and establishing anonymous internal reporting mechanisms. Framework updates must occur at least annually, with material modifications requiring publication within 30 days accompanied by justifications for changes. Beginning January 1, 2027, the Department of Technology must annually assess whether the definitions of 'frontier model,' 'large frontier developer,' and 'catastrophic risk' remain appropriate, potentially recommending legislative updates. Also by January 1, 2027, the Attorney General must report on covered employee disclosures received, and Cal OES must publish anonymized summaries of critical safety incident reports. The CalCompute consortium provisions take effect upon appropriation, with a report due by January 1, 2027, after which the consortium planning body dissolves. The phased approach allows developers to build compliance infrastructure while ensuring immediate accountability for safety practices.

Monitoring and Evaluation

SB 53 establishes multiple monitoring mechanisms creating ongoing visibility into frontier AI safety practices. The California Office of Emergency Services receives quarterly summaries of internal risk assessments and critical safety incident reports, building a government database of identified AI risks and actual incidents. These submissions are exempt from California Public Records Act disclosure, enabling candid reporting without competitive concerns, though Cal OES must publish anonymized critical incident summaries beginning January 1, 2027. The Attorney General must report annually on covered employee disclosures received, tracking whistleblower activity as an indicator of industry safety culture. The Department of Technology's annual assessment of key definitions—frontier model, large frontier developer, and catastrophic risk—provides systematic evaluation of whether regulatory thresholds remain appropriately calibrated as AI capabilities evolve. Published frontier AI frameworks and transparency reports create public records enabling external assessment by researchers, journalists, and civil society organizations. Large developers' required anonymous internal reporting mechanisms generate data on employee concerns that, while not disclosed to government, create internal monitoring capacity. The law does not establish formal performance metrics or independent evaluation requirements, relying instead on transparency through publication, government reporting streams, and whistleblower protections to surface safety issues. Market mechanisms may also contribute to monitoring as investors, customers, and partners evaluate developers' framework quality and incident history.

Penalties, Liability, and Appeals

SB 53 establishes civil penalties enforceable exclusively by the California Attorney General, with no private right of action. Violations carry penalties not to exceed $1 million per violation. Violations include failure to publish or transmit required documents (frameworks, transparency reports, incident reports), making materially false or misleading statements about catastrophic risks, failure to report critical safety incidents, and non-compliance with the developer's own published framework. Each discrete failure constitutes a separate violation, potentially generating substantial cumulative liability for persistent non-compliance. For whistleblower violations, courts may award attorney's fees to prevailing employees, creating financial incentives for private enforcement of retaliation prohibitions. In whistleblower cases, frontier developers bear the burden of proving that alleged adverse employment actions would have occurred regardless of any protected disclosure, shifting the evidentiary burden to defendants. The law provides important protections limiting liability exposure: documents may be redacted to protect trade secrets, cybersecurity vulnerabilities, public safety, or national security, provided developers retain unredacted versions for five years. Critical safety incident reports, internal use assessments, and covered employee disclosures are exempt from Public Records Act disclosure, reducing litigation and competitive risks from candid reporting. Appeals from enforcement actions follow standard California civil procedure, with defendants able to challenge penalty determinations in Superior Court and pursue appellate review. The exclusive Attorney General enforcement model concentrates prosecution decisions in a single office with statewide perspective, avoiding fragmented local enforcement.

Relationship to Other Instruments

SB 53 operates within California's expanding AI regulatory ecosystem while addressing a distinct category of advanced AI systems. California AB 2885 provides the unified AI definition applicable across California codes, while SB 53 focuses specifically on frontier models meeting computational thresholds. California AB 2013 addresses generative AI training data transparency for all developers, while SB 53's requirements apply only to large frontier developers. California SB 942 establishes AI content detection and labeling requirements based on user thresholds, complementing SB 53's safety-focused framework. Together, these laws create comprehensive coverage: AB 2885 (definition), AB 2013 (training data), SB 942 (content transparency), and SB 53 (frontier safety). The Executive Order 14110 issued by President Biden establishes federal AI safety requirements that interact with SB 53, particularly regarding national security applications exempt from state disclosure. The European Union AI Act includes obligations for providers of general-purpose AI models with systemic risks that conceptually align with SB 53's frontier model category, though specific thresholds and requirements differ. SB 53's explicit preemption of local regulations ensures statewide uniformity while preserving space for federal action. The law's catastrophic risk focus complements rather than duplicates existing product liability, tort, and criminal law, adding preventive requirements to reactive legal remedies.

International Alignment

SB 53 positions California at the forefront of international AI safety governance, establishing requirements that exceed most existing frameworks while maintaining compatibility with emerging global standards. The European Union AI Act's provisions for general-purpose AI models with systemic risk share SB 53's focus on the most capable AI systems, though the EU uses different capability assessment methodologies and institutional structures. Both frameworks emphasize risk management documentation, incident reporting, and transparency, suggesting convergence on core governance elements. The United Kingdom's approach, combining voluntary AI safety commitments with the AI Safety Institute's research agenda, differs from SB 53's mandatory compliance framework but addresses similar concerns about frontier model risks. The Bletchley Declaration signed by 28 countries in November 2023 acknowledged frontier AI risks requiring international cooperation, validating the risk categories SB 53 addresses. China's regulations on generative AI and algorithmic recommendations establish government oversight mechanisms, though with different transparency and enforcement approaches than California's framework. International technical organizations including the Partnership on AI and the Frontier Model Forum have developed governance practices that large developers may reference when implementing SB 53 requirements. The law's $500 million revenue threshold and 10^26 computing operations definition currently capture primarily U.S.-headquartered companies, though international developers serving California residents would face compliance obligations. SB 53's whistleblower protections and incident reporting requirements may influence international standards development, as other jurisdictions observe California's implementation experience with advanced AI governance.

Implementation Timeline

Sources and References