A draft bill proposed by Anthropic would limit when the US government could block an advanced artificial intelligence model, setting a threshold of a foreseeable, imminent risk of killing or seriously injuring more than 50 people — or causing more than $1 billion (£750 million) in damage — from specified catastrophic risks.
The proposal would give the US Commerce Department a central role in seeking court orders to halt qualifying AI models, raising questions about how the government could respond to emerging threats and how much influence AI developers should have over their own safety oversight.
Three people familiar with the effort told Politico that Anthropic had presented US lawmakers with a draft bill intended to regulate the safety and security of advanced AI models.
According to the report, the proposed legislation would allow the government to seek to block a model's release or use if it posed a 'foreseeable, imminent risk' of:
- Causing more than $1 billion (£750 million) in damage.
- Killing or seriously injuring more than 50 people.
The qualifying harm would have to stem from one of three specified scenarios: an AI model creating or releasing a biological weapon, evading its developer's controls or conducting a cyberattack. The thresholds concern anticipated catastrophic harm, not simply the number of people exposed to a potential danger.
High Thresholds for Federal Action
Politico reported that the draft would require independent evaluators registered with the US Commerce Department to assess AI developers and their advanced models periodically. Companies could pay these evaluators, provided their compensation was not tied to the outcome of their assessments.
If evaluators identified an 'imminent enumerated risk', they would have to report it to the Commerce Department within 72 hours of completing their assessment. The department could then ask a federal judge to suspend the model's release or use.
The 72-hour reporting period would begin after the assessment was completed, rather than when a potential threat first emerged.
The proposed restrictions on government intervention are another significant element of the draft. According to Politico, the language would prevent other federal departments, agencies and officers from suspending an AI model's development, deployment or internal use under the framework.
This would concentrate the proposed intervention process within the Commerce Department and the courts. However, the draft should not be interpreted as proving that the government would lack every other legal avenue for responding to an AI-related threat.
Who Would Set the Rules for AI Safety?
The proposal also raises questions about the independence of AI safety assessments. Although external evaluators would be responsible for examining developers and their models, the reported framework would allow the companies being assessed to pay them, subject to restrictions on outcome-linked compensation.
That arrangement creates a potential conflict-of-interest concern, even if payment alone does not establish that an evaluator would be biased. The effectiveness of the proposed oversight system would depend in part on how the evaluators were selected, how their independence was protected and how their findings were handled.
The draft would also affect the relationship between federal and state AI regulation. Reporting indicates that it would pre-empt some state requirements concerning public disclosure, incident reporting and third-party evaluations.
The proposal could therefore reshape how advanced AI systems are regulated across the US, although the precise scope of any pre-emption would need to be established from the draft itself. It should not be taken to mean that all state AI laws would be overridden.
The Wider Debate Over AI Regulation
Anthropic's proposal comes amid wider concerns about the security of increasingly capable AI systems, including their potential misuse in cyberattacks, the creation of biological weapons and the possibility that models could evade their developers' controls. These are among the risks specifically addressed by the draft.
A national framework could offer common requirements across states, potentially reducing regulatory fragmentation. However, the proposed thresholds for intervention and the limits on which federal bodies could act could also prompt questions about whether the framework would provide sufficiently broad oversight.
These are potential implications of the proposal, rather than evidence of a settled response from supporters or critics.
The central question is whether the proposed framework would provide an effective way to respond to imminent, high-impact AI risks while preserving meaningful independent scrutiny. Its thresholds, evaluator arrangements and restrictions on federal intervention are likely to be central to that debate.
The measure remains a draft proposal reported by Politico, rather than an enacted law. Its eventual impact would depend on whether lawmakers take it forward and what provisions, if any, survive the legislative process.
