Date: September 8, 2025
Core Announcement
Anthropic officially endorses SB 53, a California bill governing powerful AI systems from frontier developers. This support follows lessons from the previous attempt (SB 1047) and is driven by the need for regulation despite a preference for federal action.
"While we believe that frontier AI safety is best addressed at the federal level instead of a patchwork of state regulations, powerful AI advancements won't wait for consensus in Washington."
Key Provisions of SB 53
The bill mandates that large companies developing the most powerful AI systems must:
- Develop and publish safety frameworks describing how they manage, assess, and mitigate catastrophic risks (mass casualty incidents or substantial monetary damages).
- Release public transparency reports summarizing risk assessments and steps taken before deploying powerful new models.
- Report critical safety incidents to the state within 15 days and confidentially disclose summaries of catastrophic risk assessments for internally-deployed models.
- Provide clear whistleblower protections covering violations of these requirements and specific public health/safety dangers.
- Be publicly accountable for framework commitments, facing monetary penalties for non-compliance.
Anthropic's Rationale & Alignment
- Formalizes Existing Practices: The bill legally standardizes practices already followed by Anthropic and other frontier labs (e.g., publishing safety policies, system cards).
- Creates a Level Playing Field: Prevents a "race to the bottom" on safety disclosures by making transparency mandatory for all covered developers.
- Appropriate Scope: Focuses on large companies developing the most powerful systems, exempting startups and smaller companies to avoid unnecessary burdens.
"SB 53 implements this principle through disclosure requirements rather than the prescriptive technical mandates that plagued last year's efforts."
Areas for Future Improvement
Anthropic identifies three areas to build upon SB 53's foundation:
- Refining the Threshold: The current compute-based trigger (10^26 FLOPS) is an acceptable start but risks missing some powerful models.
- Requiring Greater Detail: Developers should provide more specifics on tests, evaluations, and mitigations undertaken.
- Enabling Regulatory Evolution: Regulations must evolve with technology; regulators need the ability to update rules as AI advances.
Conclusion & Call to Action
Anthropic commends Senator Wiener and Governor Newsom, framing SB 53 as a thoughtful, proactive approach to AI governance. The company encourages California to pass the bill and expresses willingness to work with policymakers globally.
"The question isn't whether we need AI governance—it's whether we'll develop it thoughtfully today or reactively tomorrow. SB 53 offers a solid path toward the former."
来源
暂无来源