Anthropic has proposed an AI safety bill that would establish a threshold for blocking AI model releases based on potential harm, according to TechCrunch. The proposal centers on a specific metric: a model could be halted if evidence suggests it could enable 50 or more deaths.
The bill presents a notable protection structure. It grants federal agencies authority to block a model's release, but only with a judge's permission. This requirement creates a judicial checkpoint in the approval process. Without court authorization, agencies cannot prevent deployment of AI systems, regardless of safety concerns they identify.
The proposal reflects Anthropic's approach to AI governance. By establishing numerical thresholds for harm potential, the company quantifies risk assessment for AI systems. The 50-death metric provides a concrete standard rather than vague risk categories. This specificity aims to clarify when intervention becomes necessary.
The judicial requirement represents a significant constraint on regulatory power. Federal agencies would need to present evidence to a judge and obtain court approval before blocking any model release. This process differs from direct regulatory authority that agencies typically exercise over other technologies. The structure potentially slows intervention in time-sensitive situations while ensuring legal oversight of agency decisions.
TechCrunch notes the policy contains apparent contradictions. While the proposal offers protections against harmful AI deployment, its judicial requirement effectively limits how those protections function in practice. Agencies gain blocking authority only in theory, pending successful court action.
This proposal enters ongoing debates about AI regulation. Anthropic positions itself as supporting safety-focused governance while maintaining a market-friendly framework. The bill balances concerns about AI risks with limitations on regulatory overreach, though observers dispute whether the balance adequately protects against potential harms.
The specifics of how courts would evaluate the 50-death threshold remain unclear from the available information. Implementation details about evidence standards and timeframes for judicial review also remain undefined in the current proposal outline.