HackerNews中文版

主导框架是“人类英雄，五角大楼恶棍”。我认为情况更为复杂，有三件事被低估了。 ``` 1. 法规的法律定义中包含“对手”。 ``` 美国法典第 10 编 § 3252 将“供应链风险”定义为“对手可能破坏、恶意引入不必要的功能或以其他方式颠覆”国家安全系统所带来的风险。这个词至关重要——该法规是为与中国共产党有关联的供应商和外国破坏者设计的，而不是为了与美国公司之间的合同纠纷，这些公司自愿放弃数亿美元的收入以切断与中国共产党有关联的客户。这种指定不仅在政治上是前所未有的，而且考虑到法规本身的框架，在文本上也很奇怪。 ``` 2. Anthropic 的法律挑战比报道的范围更窄。 ``` § 3252(c)(1) 包含一项不得司法审查的条款：“不得对任何行为……在政府问责局或任何联邦法院提起投标抗议进行审查。” Anthropic 的法律团队知道这一点——他们的挑战将不得不在宪法或《行政程序法》的基础上进行，而不是标准的投标抗议。这是一条更艰难的道路。“我们将在法庭上见”的说法在一定程度上误导了人们对他们实际可用的东西的看法。 ``` 3. 民主合法性问题是双向的。 ``` 大多数报道都将 Anthropic 的两次拒绝（不使用完全自主的武器，不进行大规模国内监视）视为完全正确。它们很可能确实是正确的。但“哪些人工智能系统足够可靠以做出目标决策”原则上是一个应该由民选官员和军事指挥官来决定的问题——而不是一个私营公司的首席执行官。Dario Amodei 并非民选。他的立场是可以辩护的；但它并非自动具有权威性。这与苹果/联邦调查局的 iPhone 案也不同。苹果被要求解锁现有功能。而国防部则要求 Anthropic 允许在任何现有合同中没有的新用途——是一种扩展，而不是解锁。《国防生产法》的威胁让我真正感到担忧。使用战时征用权力来强迫移除人工智能安全防护措施是一种不同类型的权力运作，没有明确的先例。一个已确认的事实让整个事情变得荒谬：据报道，美国中央司令部在宣布供应链风险指定后的几小时内，在伊朗空袭期间使用了 Claude。被指定的“供应链风险”正在实时运行国家安全行动。更难的问题是，应该用什么框架来管理私营人工智能公司以道德为由拒绝政府合同的行为，这个问题没有人认真对待。这并非因为答案显而易见，而是因为它需要思考一种不适合现有法律或政治类别的企业良知。

查看原文

The dominant frame is "Anthropic hero, Pentagon villain." I think the situation is more complicated, and three specific things are being underreported.<pre><code> 1. The statute has "adversary" in the legal definition. </code></pre> 10 U.S.C. § 3252 defines "supply chain risk" as the risk that "an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert" a national security system. That word is load-bearing — the statute was designed for CCP-linked vendors and foreign saboteurs, not contract disputes with American companies that voluntarily forewent hundreds of millions in revenue to cut off CCP-linked customers. The designation isn't just politically unprecedented; it's texturally strange given the statute's own framing.<pre><code> 2. Anthropic's court challenge is narrower than reported. </code></pre> § 3252(c)(1) includes a no-judicial-review clause: "no action...shall be subject to review in a bid protest before the Government Accountability Office or in any Federal court." Anthropic's legal team knows this — their challenge will have to be on constitutional or Administrative Procedure Act grounds, not a standard bid protest. That's a harder road. The "we'll see them in court" framing is somewhat misleading about what's actually available to them.<pre><code> 3. The democratic legitimacy question runs in both directions. </code></pre> Most coverage treats Anthropic's two refusals (no fully autonomous weapons, no mass domestic surveillance) as straightforwardly correct. They may well be. But "which AI systems are reliable enough to make targeting decisions" is, in principle, a question for elected officials and military commanders — not a private CEO. Dario Amodei wasn't elected. His position is defensible; it's not automatically authoritative.This is also different from the Apple/FBI iPhone case. Apple was asked to unlock existing capability. The DoW was asking Anthropic to allow new uses not in any existing contract — an expansion, not an unlock.The Defense Production Act threat is where I get genuinely alarmed. Using wartime conscription authority to force removal of AI safety guardrails is a different kind of power move with no clear precedent.One confirmed fact that makes the whole thing absurd:US Central Command reportedly used Claude during the Iran airstrikes hours after the supply chain risk designation was announced. The designated "supply chain risk" was running national security operations in real time.The harder question what framework should govern private AI companies refusing government contracts on ethical grounds is the one nobody's seriously engaging with. Not because the answer is obvious, but because it requires thinking about a kind of corporate conscience that doesn't fit neatly into existing legal or political categories.

Anthropic/Dow 供应链风险故事中被忽略的三个问题