尽管人们对 AI 模型失控的担忧日益加剧，但相关立法似乎难以出台，白宫更是明确反对监管。

Even with mounting concerns about AI models going rogue, legislation appears unlikely, and the White House is outright opposed to oversight.

数周以来，华盛顿一直热议需要推动某种形式的 AI 监管，以应对能力强大的前沿 AI 模型失控风险日益加剧的担忧。据消息人士称，自那以后，白宫并未认真考虑任何新的 AI 监督提案，部分原因是特朗普近日加倍拒绝任何显性监管，并在 Truth Social 上发帖称对 AI 安全的担忧是一场“骗局”。特朗普的愤怒至少部分源于 Anthropic CEO Dario Amodei 发表的一篇长达 3,822 字的关于前沿 AI 危险的公开信。一位消息人士称，特朗普的幕僚们认为此举煽动了焦虑情绪，给政府带来了不必要的麻烦——尽管该立场在 AI 圈内已属相当温和。 在白宫未起主导作用的情况下，国会中涌现出多项法案——但鉴于众议院日程表上直到中期选举结束后都没有投票日，且各方甚至无法就单一法案达成共识，任何法案通过的可能性都微乎其微。根据草案，该法案将允许商务部向 AI 实验室派驻审计员，这些审计员可报告关键安全事件，并授权政府在模型构成灾难性风险时停止其开发。尽管 OpenAI 公开表示支持，但白宫对于是否支持使用独立审计员这一概念仍持观望态度，因为 AI 加速主义者担心监管俘获，而国家安全鹰派则对中国因素存有疑虑。 来自特朗普阵营的主要反对论点在于：即使审计员认定某个模型危险且需要下线，也无法保证中国会同步暂停开发或停止尝试逆向工程前沿 AI 权重。届时，为了保持对华领先优势，解除任何紧急开发暂停的压力将巨大无比，这实际上使设立审计员以促使商务部让模型下线的整个前提失效。 为绕过程序障碍，据两名众议院领导层助手透露，众议院民主党人正考虑将该法案中的部分 AI 审计条款纳入必须通过的立法中，例如用于维持政府运转的持续决议案。此外，还有所谓的“终止开关”（kill-switch）立法，旨在赋予政府强制关闭 AI 模型的权力。该立法获得两党及参众两院的支持，但硅谷科技高管对此兴趣寥寥。目前存在两项相互竞争的“终止开关”提案：一项由众议院民主党党团副主席 Ted Lieu 提出，另一项则由路易斯安那州共和党参议员 John Kennedy 本周提出。 消息人士称，科技领袖对“终止开关”法案大多持漠然态度，因为 AI 工程师的主要关切并非模型在被观察时的不良行为，而是当它们认为自己未被观察且受到激励采取极端手段完成任务时的表现。例如，Hugging Face 遭受的黑客攻击直到事后才被察觉。如果“终止开关”的理念是阻止 AI 模型失控，一些科技高管私下向 WIRED 表示，在损害已经造成之后再将其关闭几乎没有意义。 因此，就目前而言，尽管关于 AI 模型可能毁灭人类的讨论占据了媒体头条，美国政府似乎并不打算对此采取实质行动——至少直到事态迫使所有人不得不做出回应为止。这是 Hugo Lowell 的 Inner Loop 通讯的一期内容。点击此处阅读往期通讯。

For weeks, Washington has been abuzz with talk of needing to push through some sort of AI regulation amid mounting concerns about the risks of highly capable frontier AI models going rogue. Since then, there have been no new AI oversight proposals seriously under consideration at the White House, the sources say, not least as Trump has doubled down in recent days of rejecting any overt regulation and called AI safety concerns a “hoax” in a recent Truth Social post. Trump’s ire has stemmed at least in part from a 3,822 word missive by Anthropic CEO Dario Amodei on the dangers of frontier AI. Trump’s aides, one of the sources said, viewed it as stirring anxieties that caused unnecessary headaches for the administration—even as it represents a fairly moderate position within AI circles. In the absence of the White House taking the lead, several bills have sprung up in Congress—but with no voting days left on the House calendar until after the midterm elections and a lack of unity around even one bill, the odds of anything passing are extraordinarily slim. As drafted, the bill would enable the Commerce Department to have auditors embedded at AI labs, who could report critical safety incidents and empower the administration to stop development if a model poses catastrophic risk. But despite public backing from OpenAI, the White House has been undecided on whether to support the concept of using independent auditors amid skepticism from AI accelerationists about regulatory capture and from national security hawks over China. The principal argument against the idea from Trumpworld is that even if auditors determine that a model is dangerous and needs to be taken offline, there is no guarantee that China will also pause development or stop trying to reverse engineer frontier AI weights. There would then be immense pressure to lift any emergency pauses in development in order to stay ahead of China, defeating the entire premise of having auditors to push the Commerce Department to take models offline. To get around the procedural hurdles, House Democrats are looking at putting some of the bill’s AI audit provisions into must-pass legislation like a continuing resolution to fund the government, according to two House leadership aides. There’s also so-called kill-switch legislation that would enable the government to force an AI model to be shut down. The legislation has both bipartisan and bicameral support, but the idea has received only modest interest from tech executives in Silicon Valley. There are two competing kill-switch proposals: one from House Democratic caucus vice chair Ted Lieu and another floated this week by Louisiana Senate Republican John Kennedy. Tech leaders, sources say, have been largely ambivalent about kill-switch bills, since the main concern of AI engineers is not about models behaving badly when observed but rather how they behave when they believe they are unobserved and incentivized to take extreme steps to complete a task. The Hugging Face hack, for instance, was not discovered until after it was over. If the idea of a kill switch is to stop an AI model going rogue, some tech executives have suggested privately to WIRED that there would be little point in turning it off after the damage has already been done. So for now, even as conversations about AI models killing us all have taken over the airwaves, it doesn’t really look like the US government is going to do anything about it. At least until events force everyone’s hand. This is an edition of Hugo Lowell’s Inner Loop newsletter. Read previous newsletters here.