TechCrunch reported on 20 October, citing Politico, that Anthropic is sharing a draft AI safety bill and hopes Congress will adopt it. The draft would let a court halt an AI model's release when risk is clear and near. TechCrunch says the bar is more than 50 people at risk of dying, or over $1 billion in property damage. The harm must come from a biological weapon, a model escaping its maker's control, or a cyberattack. Independent evaluators registered with the Commerce Department would judge it. They would have 72 hours to alert Commerce, which must then ask a D.C. judge to halt the release.
An Anthropic spokesperson confirmed the draft to Politico. The spokesperson said it follows the company's Advanced AI Framework, pushed with lawmakers since June. Anthropic had not answered TechCrunch's request for comment. Labs could pay the evaluators if pay does not depend on the results. TechCrunch says it is unclear how that would stop money from swaying findings. The sources do not say which lawmakers back the draft or when it might be filed.
The draft is only a proposal. It shows one AI lab's idea of how a model release could be stopped, which bears on teams that build on or buy AI models.