Abstract:
Artificial intelligence company Anthropic recently privately submitted a draft artificial intelligence safety bill to U.S. lawmakers in an attempt to establish a federal-level regulatory framework for advanced AI models. According to the proposal, if an AI model is determined to have a foreseeable and imminent catastrophic risk, which may cause more than 50 deaths or serious injuries, or cause more than $1 billion in property damage, its release and use may be blocked.
According to the disclosed information, the bill does not target all artificial intelligence applications, but focuses on the most advanced "cutting-edge models." The proposal sets a clear threshold for intervention, and only in extreme risk situations can regulators initiate legal proceedings to restrict model deployment.
The risks listed in the draft mainly focus on three areas: assisting in the creation or release of biological weapons, models escaping from the control of developers, and large-scale cyber attacks participated in or promoted by models. Only when risks arise from these specific scenarios will regulatory mechanisms be triggered.
It is worth noting that the risk standards proposed by Anthropic are similar to Senate Bill 53 previously passed by California, USA. The latter also defines situations that may cause large numbers of casualties or huge economic losses as “catastrophic risks.”
As for the specific implementation mechanism, the bill proposes to establish an independent evaluator system registered and managed by the U.S. Department of Commerce. These assessment agencies or experts are responsible for conducting security assessments of cutting-edge AI models and their development companies, and determining whether relevant models pose significant risks.
According to the draft arrangement, even if the evaluation agency is paid by the AI company, this partnership is still allowed as long as the payment is not directly related to the evaluation results. However, this arrangement has also caused doubts among some industry insiders, who believe that the independence and objectivity of the evaluation agency may face challenges.
If evaluators determine that a model presents an imminent catastrophic risk, they must report it to the U.S. Department of Commerce within 72 hours of completing the evaluation. The Commerce Department could then petition a federal court in Washington, D.C., to ask a judge to grant a moratorium on the release or deployment of the model.
Unlike some regulatory plans, Anthropic's proposal specifically emphasizes that administrative agencies do not have the power to unilaterally halt models. Under the draft design, only a federal judge could ultimately issue a moratorium. In other words, no government department can directly ban an AI model from going online or using it, bypassing judicial procedures.
Anthropic has not publicly detailed the final version of the bill, but a company spokesperson has confirmed that the draft does exist and said that its core ideas stem from the "Advanced Artificial Intelligence Framework" proposed earlier this year. The framework advocates more rigorous testing, greater transparency, and the introduction of independent third-party evaluation mechanisms for the most powerful AI systems.
As artificial intelligence capabilities continue to improve, the debate surrounding AI safety issues among the U.S. technology industry and regulatory agencies is also heating up. Supporters believe that the plan proposed by Anthropic provides clear procedures and legal basis for government intervention in high-risk AI; while critics believe that the excessive intervention threshold and reliance on industry self-assessment mechanisms may make it difficult to fully respond to more complex artificial intelligence risks in the future.
Currently, this proposal is still in the discussion stage, has not yet been formally submitted to the U.S. Congress for consideration, and does not have legal effect. However, as one of the first federal regulatory plans proactively proposed by large AI companies, its content is expected to have an important impact on future discussions on artificial intelligence legislation in the United States.
Comments