On October 16, 2026, according to Reuters and Politico, citing three people familiar with the matter, representatives of Anthropic quietly circulated a draft frontier AI safety bill among U.S. Capitol Hill staff. The core provision is just one sentence: for the government to ban a frontier model, it must prove to a court that the model would cause more than $1 billion in losses or at least 50 deaths or serious injuries. Below that line, any federal department, agency, or even official is explicitly barred from intervening in a model’s development, deployment, or internal use.
This is not a regulatory text commissioned by the government. It is an AI company valued at tens of billions of dollars that has itself designed, for its own industry, the standard for “when it can be stopped.”
The Draft’s Mechanism Design
The draft is not structured as a simple “high threshold, then let it through” approach; it establishes a three-tier mechanism.
The first tier is an independent evaluation system. Frontier AI labs would have to undergo periodic review by independent third-party evaluation bodies registered with the Department of Commerce. The companies being evaluated would pay the evaluation fees, with the only restriction being that “compensation must not be tied to the evaluation conclusions.” This arrangement resembles the pre-financial-crisis model in which rating agencies were paid by the entities they rated.
The second tier is a 72-hour reporting obligation. If an evaluator determines that a model poses an “imminent enumerated risk,” it must report to the Department of Commerce within 72 hours after completing the evaluation. The Commerce Department could then apply to a federal district court for a suspension order, but only a federal judge located in the District of Columbia would have the authority to issue that order.
The third and most critical tier is a concentration of authority: the draft explicitly prohibits any other federal department, agency, or official from unilaterally halting the development or deployment of an AI model. In other words, situations like this summer, when export controls forced two Anthropic models to be shelved, would be foreclosed at the legal level under this framework.
An Anthropic spokesperson confirmed to Reuters that the draft is a legislative translation of the company’s “Advanced AI Framework,” released earlier this year, and said the company has been lobbying federal and state legislatures since June.
Where the $1 Billion Threshold Comes From
The number did not come out of thin air. According to Benzinga, the threshold language in the draft closely matches provisions in California’s SB 53, which was signed into California law last year. By anchoring the federal proposal in an existing state-law precedent, Anthropic gives the standard a greater appearance of legitimacy on the grounds of “aligning with existing law,” while also endorsing a uniform national standard.
But $1 billion as a quantitative threshold for triggering government intervention is itself an extremely high starting point. Before an AI system actually causes quantifiable, large-scale economic loss, any preventive intervention would struggle to clear this bar. The framework responds only to disasters that have already happened, rather than intervening before a disaster takes shape.
Supporters see this as a pragmatic consideration: it provides an enforceable legal standard and prevents arbitrary government intervention in the absence of a clear basis. Critics, however, point out that the line sits precisely where the vast majority of model controversies can never reach.
The Structural Dilemma of “Referee and Player”
At the heart of this controversy is not whether the threshold is high or low, but who has the power to set it.
Anthropic is one of the world’s top three frontier AI developers, and its Claude series models would fall directly under this framework. Its drafting the rules for “under what circumstances a model can be banned” is structurally identical to an airline drafting the airworthiness standards for “under what circumstances a plane must be grounded.” The tension between commercial interests and regulatory standards exists naturally without any subjective malice.
This is not the first time such a phenomenon has appeared in the tech industry. Internet platforms have long helped shape the interpretive space around Section 230 of the Communications Decency Act; financial institutions directly participated in negotiations over the details of the Dodd-Frank Act after the 2008 crisis; and the tobacco industry spent decades hiring lobbyists to influence standards for determining the harms of smoking. The difference is that AI’s risk window is extremely narrow—the time span from model release to possible catastrophic use may be measured in days, not years. In such a field, letting the regulated party set the pace of legislation may carry costs far greater than in the past.
Another provision in the draft is that independent evaluators would be paid by the companies being evaluated, and the operation of the entire system would not involve any government-initiated substantive review capacity. This means that under this framework, the federal government would in effect be a passive recipient that can act only after waiting for a “passive report,” rather than an active regulatory force.
A Crowded Legislative Race
Anthropic’s draft is not the only AI legislative proposal on Capitol Hill. According to Interesting Engineering, Representatives Trahan and Obernolte have advanced a competing “FRONTIER Act”; Senate Majority Leader Thune is working with Klobuchar and Cruz on legislation addressing catastrophic risk. Meanwhile, Representative Casar and Senator Sanders advocate a direct pause on advanced AI development.
This landscape means there is uncertainty over whether Anthropic’s draft can become the final legislative basis. But its strategic significance lies not entirely in passage, but in defining the boundaries of the debate. Once $1 billion and 50 deaths or serious injuries become the reference point in public debate, other, stricter standards will have to start from an obligation to explain “why the threshold should be lowered,” rather than Anthropic having to explain “why it is set so high.”
The draft also contains a potential political bombshell: it would preempt, at the federal level, the room for states to adopt stricter AI safety rules. States such as California and New York, which already have relatively strong local regulations, may find their legislative authority encroached upon by this “uniform federal standard”—a standard drafted precisely by the largest AI developers.
Independent Judgment
The real problem with Anthropic’s move is not whether it is “bad,” but that the manner of participation itself lacks transparency. The draft was not submitted through a public proposal process; it was “quietly circulated” among Capitol Hill staff and entered public view only after Politico disclosed it. This way of operating contrasts with the principle of “AI safety transparency” that Anthropic repeatedly emphasizes in its official statements.
Supporting an enforceable legislative framework has value—vague safety principles cannot constrain anyone. But there is a substantive difference between “enforceable” and “favorable to the industry,” and the boundary between the two in this draft needs careful scrutiny by independent regulators, legal scholars, and legislators, rather than being defined by the industry itself.
At this moment, a more fundamental question is worth pressing: when the company most capable of understanding frontier AI risks is also the company with the strongest incentive to lower the intervention threshold, what institutional mechanisms can we build to truly separate these two roles? Anthropic’s draft does not answer this question—it structurally avoids it.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接