Anthropic Endorsement of California SB 53

Anthropic has officially endorsed California Senate Bill 53 (SB 53), a legislative effort to govern the development of powerful AI systems. The bill aims to establish a mandatory transparency and disclosure framework for frontier AI developers to mitigate catastrophic risks without imposing the prescriptive technical mandates seen in previous regulatory attempts.

Core Requirements of SB 53

SB 53 shifts the regulatory focus toward a "trust but verify" approach, requiring large companies developing the most powerful AI systems to adhere to the following mandates:

Safety Frameworks: Developers must create and publish safety frameworks detailing how they manage, assess, and mitigate catastrophic risks—defined as risks that could materially contribute to substantial monetary damages or mass casualty incidents.
Transparency Reports: Before deploying new powerful models, companies must release public reports summarizing their catastrophic risk assessments and the steps taken to fulfill their safety frameworks.
Incident Reporting: Critical safety incidents must be reported to the state within 15 days. Additionally, developers must confidentially disclose summaries of assessments regarding the potential for catastrophic risk from internally-deployed models.
Whistleblower Protections: The bill provides legal protections for employees who report violations of these requirements or specific, substantial dangers to public health and safety stemming from catastrophic risk.
Public Accountability: Companies are legally held accountable for the commitments made in their frameworks, with monetary penalties for non-compliance.

Scope and Industry Alignment

SB 53 targets large companies developing the most powerful AI systems, providing exemptions for startups and smaller companies to avoid unnecessary regulatory burdens.

Anthropic notes that these requirements formalize practices already adopted by several frontier labs. For example, Anthropic utilizes a Responsible Scaling Policy and publishes system cards to document model capabilities and limitations. Similar frameworks have been been adopted by Google DeepMind, OpenAI, and Microsoft. By making these disclosures mandatory, SB 53 creates a level playing field, preventing labs from reducing safety and disclosure programs to gain a competitive advantage.

Technical Thresholds and Future Improvements

Anthropic identifies SB 53 as a strong foundation but suggests three areas for future regulatory evolution:

Training Compute Thresholds

The bill currently identifies regulated AI systems based on the amount of computing power used for training, with a threshold set at 10^26 FLOPS. Anthropic notes that while this is an acceptable starting point, there is a risk that some powerful models may fall outside this scope.

Detail of Evaluations

Anthropic advocates for developers to provide more granular detail regarding the tests, evaluations, and mitigations they perform. They highlight their own work with the Frontier Model Forum as an example of how sharing safety research and red team testing strengthens the overall safety ecosystem.

Adaptive Regulation

Because AI technology advances rapidly, Anthropic argues that regulators must have the ability to update rules dynamically to maintain a balance between safety and innovation.

Regulatory Context

While Anthropic maintains that frontier AI safety is ideally addressed at the federal level to avoid a patchwork of state regulations, the company supports SB 53 because powerful AI advancements are progressing faster than federal consensus in Washington.