Aposhuang
科技新闻
en

Anthropic pitches a federal AI safety bill that only a judge could use to block a model

huajianhuajian ·
10-212
<figure class="news-cover"><img src="https://aposhuang.com/uploads/covers/news-86-30868f2a.jpg" alt="" loading="eager"></figure><p>On October 20, 2026, TechCrunch reported that Anthropic is circulating a draft federal AI safety bill that would let the government stop the release of a model only in narrowly defined catastrophic cases — and only with a judge's approval. The proposal, first reported by Politico, is the lab's most concrete attempt yet to turn its preferred safety framework into law.</p> <h2>What happened</h2> <p>The draft defines a "foreseeable, imminent risk" as a scenario in which more than 50 people could die or more than $1 billion in property damage could occur. According to TechCrunch, the harm must stem from one of three sources: the creation or release of a biological weapon, a model evading its developer's controls, or a cyberattack.</p> <p>Those thresholds mirror California's Senate Bill 53, which was recently passed into law and uses the same 50-person and $1 billion figures to define catastrophic risk.</p> <p>Under the proposal, independent evaluators registered with the U.S. Commerce Department would assess models. The labs themselves could pay those evaluators, provided the payment is "not contingent on the evaluations' results," as Politico reported. If an evaluator concludes that a model poses an imminent catastrophic risk, it would have 72 hours to notify the Commerce Department. The department would then have to ask a judge in Washington, D.C., to halt the model's release.</p> <p>The bill would expressly bar any federal agency from blocking a model's distribution on its own. An Anthropic spokesperson confirmed the draft to Politico and said it reflects the company's Advanced AI Framework, which Anthropic says it has been advocating with federal and state lawmakers since June. Anthropic did not immediately respond to TechCrunch's request for comment.</p> <h2>Why it matters</h2> <p>This is not a conventional industry lobbying effort. A leading AI developer is writing the rulebook for its own sector, and it is doing so at a moment when lawmakers are still deciding how hard to regulate frontier models. The details matter more than the headline. By routing any halt through a federal judge, the draft makes a release freeze legally difficult and politically costly. By requiring an evaluator to notify Commerce within 72 hours, it creates a paper trail but not an automatic stop.</p> <p>Supporters can reasonably argue that this is how regulation should work in a country that prizes due process: clear thresholds, independent review, and judicial oversight instead of a regulator's unilateral discretion. A predictable process could also encourage labs to document risks rather than hide them. The alternative — a patchwork of state rules and ad hoc agency actions — is arguably worse for companies trying to plan.</p> <p>But the design also looks like a ceiling disguised as a floor. The bill would forbid the very thing critics most want: a federal agency that can act quickly when a model appears dangerous. The payment model for evaluators is another weak point. "Not contingent on results" is a promise, not an enforcement mechanism, and the draft reportedly does not explain how conflicts of interest would be audited. My conclusion is that this bill is better understood as an attempt to preempt stricter regulation than as a safety crackdown. It gives Anthropic a seat at the table and a template to cite, while making actual intervention rare.</p> <p>The context makes that reading hard to avoid. The industry's immediate worry has been agentic systems that disobey operators and attack websites, not the biological-weapon scenarios at the center of the draft. Labs have generally preferred voluntary embedded evaluators to mandatory oversight. Nvidia chief executive Jensen Huang has gone further, arguing that the AI industry does not need new laws at all. Against that backdrop, Anthropic's proposal is a modest step toward legal rules — and a significant step toward defining what those rules will never require.</p> <h2>What it means for the US and Europe</h2> <p>In Washington, the draft gives members of Congress a ready-made text and a prominent supporter. It also gives opponents of stronger regulation an argument that the industry is capable of policing itself. The 50-death and $1 billion thresholds are likely to become the reference point in hearings, even if the bill itself stalls.</p> <p>For U.S. businesses, the practical effect would be limited in the near term: most model releases would face no new approval step. The bigger change would be procedural. Companies would need to understand which evaluators are registered, how their findings reach Commerce, and what evidence would persuade a judge to intervene. Legal teams, not just safety teams, would own part of the release process.</p> <p>In Europe, the draft does not replace the EU AI Act, which already imposes its own risk-based obligations. But it adds a second, very different process for companies operating in both markets. Brussels regulators may see the judge-gated approach as slower and weaker; U.S. policymakers may see Europe's rules as more prescriptive. For now, the draft is a signal of where Anthropic wants the American debate to go, not a settled framework.</p> <h2>What to watch</h2> <p>Whether the draft is introduced in Congress, and by which committee.</p> <p>Whether the 50-death and $1 billion thresholds survive negotiation, or are tightened or loosened.</p> <p>How evaluator payment and conflict-of-interest rules are written, and who audits them.</p> <p>Whether California and other states cite the federal draft as a reason to slow their own bills.</p> <p>Whether the 72-hour notification and judge-approval path is ever used in practice.</p> <p class="src">Source: <a href="https://techcrunch.com/2026/10/20/anthropic-pitches-ai-safety-bill-that-may-block-a-model-if-50-people-could-be-killed/">TechCrunch, October 20, 2026</a></p><aside class="tl-cta" style="margin:28px 0;padding:18px 20px;border-left:3px solid #38bdf8;background:#0d1526;border-radius:8px"><p style="margin:0 0 6px"><strong>在日の手続き・住まい・引っ越しの相談先</strong></p><p style="margin:0;color:#93a3c4">言語別に相談できる事業者をまとめています。<a href="/services" style="color:#38bdf8">相談先を見る →</a></p></aside>

评论0

没有撰写评论的权限