Under Anthropic's draft bill, the US government could block an AI model's release or use only if it poses a foreseeable, imminent risk of killing or seriously injuring more than 50 people or causing over $1 billion in damages