Anthropic drafts federal AI safety bill allowing court-ordered model halts
Anthropic has privately submitted a draft AI safety bill to U.S. lawmakers that would create a federal framework for the most advanced “frontier” models, according to a cnBeta report; an Anthropic spokesperson confirmed the draft exists. The proposal would let a federal judge in Washington, D.C., pause a model’s release or deployment if evaluators find a foreseeable and imminent catastrophic risk — such as more than 50 deaths or serious injuries, or more than $1 billion in property loss — tied to biological weapons, loss of developer control or large-scale cyberattacks. It envisions independent evaluators registered by the Commerce Department, a 72-hour reporting requirement, and allows AI companies to pay evaluators as long as payment is not directly tied to results; the draft has not been formally submitted to Congress and has no legal effect.
Anthropic’s spokesperson said the draft stems from its earlier Advanced AI Framework, making this one of the first federal regulatory proposals put forward by a major AI company — though it remains a discussion draft that has not been formally submitted to Congress.
人工智能公司Anthropic近日向美国国会议员私下提交了一份人工智能安全法案草案,试图为先进AI模型建立联邦层面的监管框架。根据提案内容,如果某个AI模型被认定存在可预见且迫在眉睫的灾难性风险,可能导致超过50人死亡或重伤,或者造成超过10亿美元财产损失,其发布和使用有可能被阻止。
根据披露信息,该法案并非针对所有人工智能应用,而是重点关注最先进的“前沿模型”。提案设定了明确的干预门槛,只有在极端风险情况下,监管机构才可启动法律程序限制模型部署。
草案列出的风险主要集中于三个领域:协助制造或释放生物武器、模型脱离开发者控制,以及由模型参与或推动的大规模网络攻击。只有当风险来源于这些特定场景时,监管机制才会被触发。
值得注意的是,Anthropic提出的风险标准与美国加利福尼亚州此前通过的《参议院第53号法案》较为相似。后者同样将可能造成大量人员伤亡或巨额经济损失的情况定义为“灾难性风险”。
#Anthropic#U.S. Congress#U.S. Department of Commerce#California Senate Bill 53#Anthropic Advanced AI Framework