California AI Transparency Act (SB 53)

This article establishes the core definitions that determine who and what is regulated by SB 53. A 'Frontier model' is defined as a foundation model trained with over 10^26 computing operations, including fine-tuning. A 'Frontier developer' is anyone who trained such a model. The most stringent requirements apply to 'Large frontier developers,' which are frontier developers with over $500 million in annual gross revenue.

The law also precisely defines 'Catastrophic risk' as a foreseeable, material risk of causing death or serious injury to 50 or more people, or property damage exceeding $1 billion, stemming from specific AI misuse scenarios like CBRN weapon assistance, cyberattacks without human oversight, or model evasion. 'Critical safety incident' covers events like unauthorized access to model weights or the materialization of catastrophic risks. 'Covered employee' refers to staff involved in assessing or managing critical safety risks, who are protected by whistleblower provisions.

Large frontier developers are required to establish and publicly publish a comprehensive 'Frontier AI Framework' on their websites. This framework must detail their approach to identifying, assessing, and mitigating catastrophic risks associated with their AI models. It needs to incorporate national and international standards and industry best practices.

Key components of this framework include defining and assessing capability thresholds that could lead to catastrophic risks, documenting specific mitigation measures, addressing the use of independent third-party evaluators for risk assessments, and outlining cybersecurity measures to protect unreleased model weights from unauthorized access. Developers must also establish internal governance structures to ensure the framework's effective implementation. These frameworks must be reviewed and updated at least annually, with material modifications published within 30 days along with justifications.

Large frontier developers are obligated to submit summaries of their internal catastrophic risk assessments to the California Office of Emergency Services (Cal OES) on a quarterly basis. This reporting mechanism allows a state emergency management agency to monitor potential AI-related threats and integrate them into broader emergency preparedness efforts.

These submitted summaries are exempt from public disclosure under the California Public Records Act, which aims to encourage candid and comprehensive reporting from developers without concerns about competitive disadvantages or litigation risks. The specific schedule for these quarterly reports will be coordinated directly with Cal OES.

Before or concurrently with the deployment of any new frontier model, large frontier developers must publish transparency reports. These reports are intended to provide the public with clear information about the AI system being released.

Each report must detail the model's capabilities, the specific safety measures that have been implemented to mitigate risks, and the intended uses of the model. This requirement aims to foster greater public understanding and accountability for the introduction of powerful new AI systems, ensuring that stakeholders are informed about their potential impacts and the safeguards in place.

Large frontier developers are required to report 'critical safety incidents' to the California Office of Emergency Services (Cal OES). The standard reporting timeline is within 15 days of discovering such an incident. However, if an incident poses an imminent risk of death or serious injury, an expedited report must be made within 24 hours of discovery.

Critical safety incidents include events like unauthorized access to model weights that could lead to harm, the actual materialization of a catastrophic risk, loss of control over the AI model, or deceptive model behavior that subverts developer controls. These reports are crucial for enabling rapid government response and monitoring of actual AI-related threats. Similar to risk assessment summaries, these incident reports are exempt from public disclosure.

SB 53 includes robust protections for 'covered employees' who report concerns related to AI safety. It explicitly prohibits retaliation against employees who disclose information about potential catastrophic risks or critical safety incidents. This protection applies whether the disclosure is made to government authorities, company management, or other employees with investigative authority.

To facilitate such reporting, developers are required to establish anonymous internal reporting mechanisms. Furthermore, companies must post notices in the workplace informing employees of their rights under these whistleblower provisions. In cases of alleged retaliation, the burden of proof shifts to the developer to demonstrate that any adverse employment action would have occurred regardless of the protected disclosure, strengthening employee protections.

The California Attorney General holds exclusive authority for enforcing SB 53, meaning no private right of action exists for individuals to sue under this law. Violations can lead to significant civil penalties, with fines not exceeding $1 million per violation. Each distinct failure to comply, such as failing to publish a required document or making a materially false statement about catastrophic risks, constitutes a separate violation, potentially leading to substantial cumulative penalties.

Specific violations include failure to publish or transmit required documents (frameworks, transparency reports, incident reports), making materially false or misleading statements regarding catastrophic risks, failure to report critical safety incidents, and non-compliance with the developer's own published framework. In whistleblower cases, courts may award attorney's fees to prevailing employees, providing a financial incentive for employees to pursue valid claims of retaliation. The law also includes protections for developers, allowing redaction of trade secrets, cybersecurity vulnerabilities, or national security information from disclosures, provided unredacted versions are retained for five years.

To ensure the regulation remains relevant and effective in a rapidly evolving technological landscape, SB 53 mandates that the California Department of Technology conduct an annual assessment. Starting January 1, 2027, the Department must evaluate whether the key definitions—'frontier model,' 'large frontier developer,' and 'catastrophic risk'—continue to be appropriate as AI capabilities advance.

This provision creates a built-in mechanism for regulatory adaptation, allowing the state to adjust the scope and focus of the law without requiring new legislative action each time AI technology progresses significantly. The Department's assessments may lead to recommendations for legislative updates if the definitions are found to be outdated or no longer accurately capture the intended scope of regulation.

SB 53 includes a preemption clause that prevents local regulations from creating a fragmented regulatory environment for advanced AI safety. Specifically, the act preempts any local regulations adopted after January 1, 2025, that specifically aim to regulate frontier developers' management of catastrophic risks.

This provision ensures statewide uniformity in the governance of frontier AI safety, meaning that developers operating across California will be subject to a single, consistent set of state-level requirements rather than a patchwork of potentially conflicting local ordinances. This simplifies compliance for developers and provides clarity regarding the applicable regulatory framework.

Beyond direct regulatory requirements, SB 53 also directs the development of the 'CalCompute' public cloud computing consortium. The primary objective of this initiative is to democratize access to essential AI research resources, particularly high-performance computing capabilities.

This aims to address concerns about the concentration of powerful computing resources among a few large entities, fostering a more diverse and accessible AI research ecosystem. While the consortium's full implementation is contingent on appropriations, a report on its planning and progress is due by January 1, 2027, after which the initial planning body will dissolve.