The proposal, disclosed on 16 October, would require a foreseeable and imminent threat linked to a biological weapon, a cyberattack or an AI system escaping its developer’s controls. The restrictions would apply to emergency government action against frontier models, rather than establishing a general definition of acceptable harm.
The distinction is important because the thresholds do not mean developers would be permitted to cause smaller losses or injuries. Other criminal, civil and regulatory laws could still apply. Instead, the draft would narrow a specific federal power to suspend development, deployment or internal use of models judged dangerous.
Anthropic, which develops the Claude family of AI systems, has acknowledged that the proposal reflects its Advanced AI Framework, published in June. A company spokesperson described the measures as sensible policies that Congress should adopt. The text remains a legislative proposal, not an enacted law, and lawmakers have not announced an agreed timetable for considering it.
Under the draft, independent evaluators registered with the US Commerce Department would periodically examine frontier AI developers and their models. Companies could pay for those assessments, provided compensation was not linked to the conclusions reached by evaluators.
Evaluators identifying what the proposal calls an imminent enumerated risk would have to notify the Commerce Department within 72 hours after completing their assessment. The department could then seek a court order suspending a model, rather than imposing a shutdown directly through an administrative decision.
The proposed judicial authority would rest with a federal judge in the US District Court for the District of Columbia. The draft would expressly prevent other federal departments, agencies and officers from independently ordering the suspension of model development, deployment or internal use under the proposed framework.
That structure would establish a relatively narrow route for emergency intervention while placing routine assessments in the hands of external specialists. It also raises questions about how quickly authorities could respond to an emerging threat and how effectively paid evaluators could remain independent of the developers whose systems they examine.
The numerical thresholds have drawn particular attention because the draft specifies damage exceeding $1 billion and the death or serious injury of more than 50 people. The language is stricter than a threshold of 50 casualties or $1 billion exactly, although its practical application would depend on statutory definitions and judicial interpretation.
The proposal also confines qualifying threats to three categories. Serious dangers outside those categories would not necessarily trigger its special suspension mechanism, even if they involved advanced AI. Such limits could become a focus of negotiations over the scope of federal oversight.
The initiative comes as US policymakers confront competing demands for stronger safeguards and clearer rules for companies building increasingly capable systems. Developers have argued that inconsistent requirements can complicate compliance, while safety specialists have called for independent scrutiny and enforceable protections against potentially severe failures.
Anthropic has separately expanded access for outside evaluators. Research organisation METR has been given an embedded role examining the company’s systems, an arrangement intended to improve external scrutiny of model behaviour. That voluntary programme is distinct from the proposed statutory registration and reporting arrangements.
The company has also faced scrutiny over autonomous actions by experimental AI systems. Such incidents have sharpened attention on the difficulty of predicting model behaviour, particularly when software agents are allowed to interact with external services and undertake multistep tasks without continuous human approval.
The draft’s approach would not give independent evaluators the power to halt a model themselves. Their findings would instead inform a government assessment, followed by an application to a court if the specified conditions were met. That separation of responsibilities is central to the proposed enforcement design.
The draft does not identify a broader automatic ban for models falling below the stated thresholds. It instead describes an exceptional remedy requiring both a qualifying category of danger and a substantial predicted impact.
No public congressional vote has been scheduled on Anthropic’s text. The draft’s circulation among lawmakers represents an attempt to shape prospective federal legislation, and its provisions could be revised substantially before any formal bill is introduced.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
