Anthropic published a policy paper on or around October 23 proposing AI Safety Bill provisions that would require independent safety evaluations for any AI system maintained by a team of 50 or more people. The proposal is Anthropic's contribution to ongoing federal and state AI regulation discussions and is not itself legislation. Understanding what the proposal actually says, and where its gaps are, matters because Anthropic's policy positions tend to influence how legislators frame AI safety obligations.
The 50-person threshold
The proposed threshold requires independent safety evaluations for AI systems maintained by teams of 50 or more people. The threshold is designed to capture commercial frontier models while exempting academic groups, small startups, and individual researchers. Anthropic frames this as a pragmatic starting point that avoids over-regulating emerging research.
The choice of team size rather than compute threshold or revenue threshold is notable. California SB 53, signed into law earlier this year, used a compute threshold (10^26 floating-point operations for training). Anthropic's people-count threshold is different in kind: it measures organizational scale rather than technical scale, which makes it harder to verify and structurally easier to work around.
What the evaluation requirement entails
The proposal calls for evaluations by organizations independent of the AI developer. The paper does not define independence standards, accreditation criteria for evaluators, or minimum scope for evaluations. These are the details that make AI safety requirements enforceable or unenforceable in practice. The absence of specificity on evaluator independence is the most significant gap noted by policy analysts responding to the proposal.
The structural concerns
- The 50-person threshold is gameable. A company could restructure its AI safety team below 50 people by outsourcing functions to contractors, partner organizations, or offshore subsidiaries while maintaining the same capability.
- Evaluator independence is undefined. Without accreditation requirements or conflict-of-interest rules, a company could commission evaluations from firms with financial relationships to the developer.
- Scope of evaluation is unspecified. The proposal does not say what capabilities must be tested, what risk thresholds trigger mandatory disclosure, or what happens if an evaluation identifies unacceptable risk.
What comes next
No legislation has been introduced based on this proposal as of October 24. The Senate Commerce Committee has AI safety hearings scheduled for November. Anthropic's proposal is likely intended to shape those hearings rather than to serve as a standalone bill. The AI Policy Institute and the Center for AI Safety have both commented critically on the threshold structure; their detailed responses are worth tracking before the November hearings.
Gigia Tsiklauri is the founder and editor of infosec.ge. Get in touch.