Anthropic proposes AI safety bill with risk thresholds for model blocking
In brief
- Anthropic proposes a bill to block AI models deemed risky based on strict thresholds (50 deaths or $1B in damage).
- The draft requires independent evaluators, paid by labs but without result-based incentives, and judicial approval to suspend a model.
- This approach, inspired by a California law, goes further by centralizing decisions via judges, despite industry resistance to strong regulation.
Anthropic has pitched a draft AI safety bill to U.S. lawmakers, setting specific thresholds for blocking a model's release. According to Politico, the proposal aims to prohibit distribution if an "imminent and foreseeable risk" threatens over 50 lives or causes over $1 billion in property damage, including scenarios like biological weapons, control evasion, or cyberattacks.
The bill introduces independent evaluators, registered with the U.S. Department of Commerce, to assess risks. These evaluators could be paid by AI labs, provided payments are not tied to evaluation outcomes. If a risk is identified, the Commerce Department would have to request a Washington, D.C., judge to halt the model's release, removing direct agency discretion.
This framework partly mirrors California's recent SB 53 law but stands out by relying on judicial decisions. It emerges as labs, despite warnings about existential risks, resist mandatory oversight, favoring voluntary systems like embedded evaluators.
What it changes for you
- Public SectorGovernments may need to implement an AI risk evaluation system with legal thresholds, certified evaluators, and judicial oversight.
This summary gives you the essentials. The full article, its figures and quotes are on AI News & Artificial Intelligence | TechCrunch.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error