Anthropic announced on September 8, 2025, that it supports California Senate Bill 53, a proposal to regulate transparency and safety obligations for the most powerful AI systems produced by frontier developers. The company said it has long advocated for thoughtful AI regulation and that its endorsement follows careful consideration of lessons from California’s earlier regulatory attempt, SB 1047.
Why the law matters
Anthropic maintains that frontier AI safety is best handled at the federal level rather than through a patchwork of state rules, but acknowledges that technological progress will not wait for federal consensus. Governor Gavin Newsom convened the Joint California Policy Working Group, a panel of academics and industry experts, which recommended a “trust but verify” approach. Senator Scott Wiener’s SB 53 seeks to implement that principle by imposing disclosure obligations rather than prescriptive technical mandates.
Key provisions of SB 53
Under the bill, large companies developing the most powerful AI systems would be required to:
- Develop and publish safety frameworks explaining how they manage, assess, and mitigate catastrophic risks—those that could foreseeably and materially contribute to mass casualties or substantial monetary damages.
- Release public transparency reports summarizing catastrophic risk assessments and the steps taken under their frameworks prior to deploying powerful new models.
- Report critical safety incidents to the state within 15 days and confidentially provide summaries of any assessments of catastrophic risk potential from internally deployed models.
- Provide clear whistleblower protections covering violations of these requirements as well as specific and substantial dangers to public health or safety arising from catastrophic risks.
- Be publicly accountable for commitments made in their frameworks, with monetary penalties for noncompliance.
The bill targets large developers of the most powerful systems while exempting startups and smaller firms that are less likely to build such models, reducing unnecessary regulatory burden on smaller actors.
Relation to existing industry practice
Anthropic says SB 53 would formalize practices already followed by it and many other frontier labs. Anthropic publishes a Responsible Scaling Policy that outlines how it evaluates and mitigates risks as models gain capability, and it issues system cards documenting model capabilities and limitations. Other frontier organizations — Google DeepMind, OpenAI, and Microsoft — have adopted similar transparency and safety practices while competing at the frontier.
By making disclosure mandatory for covered models, the bill aims to create a level playing field so developers can compete without being incentivized to weaken safety or transparency programs.
Areas for improvement noted by Anthropic
Anthropic welcomes SB 53 as a strong foundation but encourages policymakers to refine certain aspects:
- The bill currently determines which systems fall under regulation based on computing used in training (FLOPS). The present threshold, 10^26 FLOPS, is an acceptable starting point, but Anthropic warns some powerful models might still fall outside the threshold.
- Developers should be required to provide more detailed information about the tests, evaluations, and mitigations they perform. Publishing safety research, red team testing documentation, and deployment rationales—as pursued through industry coordination such as the Frontier Model Forum—strengthens safety.
- Regulations should be able to evolve as AI technology advances; rulemakers need mechanisms to update requirements to maintain the right balance between safety and innovation.
Closing
Anthropic praised Senator Scott Wiener and Governor Gavin Newsom for leadership on responsible AI governance. The company argued that the question is not whether to have AI governance, but whether it will be developed thoughtfully now or reactively later. Anthropic urged California to pass SB 53 and said it looks forward to working with policymakers in Washington and internationally to develop comprehensive approaches that protect public interests while sustaining American leadership in AI.
