Anthropic drafts AI safety bill setting specific thresholds for federal intervention

Artificial intelligence startup Anthropic has circulated a draft safety bill to United States lawmakers, TechCrunch reported. Politico first reported that the company pitched the legislative proposal to Congress to establish formal rules for government regulation of the industry. TechCrunch and UA.NEWS both report that the draft limits when federal authorities can restrict a model. [1][2]

Under the proposal, federal agencies are forbidden from blocking the release of an AI model without permission from a judge, according to TechCrunch. The draft legislation specifies scenarios where model distribution could be halted. Mezha reported that these conditions include situations where a model causes at least 50 deaths or results in 1 billion dollars in damages. [1][6]

Reports from Truthout, Futurism, and Cybernews confirm that the proposed framework requires a minimum threshold of 50 potential human casualties before government authorities can intervene to stop an AI model. Futurism noted that the company has been lobbying government officials on the draft text. [3][4][5]

Debates over artificial intelligence safety, alignment, ethics, and severe future risks have accelerated during the current development boom. Anthropic operates with an independent Long-Term Benefit Trust to oversee its public benefit purpose and uses a training methodology known as Constitutional AI to shape model behavior through explicit natural-language principles. [7][8]