内容来源：https://www.wired.com/story/whatever-ai-safety-looks-like-its-not-this/

内容总结：

1966年，美国高速公路事故导致近5.1万人死亡。大量证据表明，如果当时已问世数年的现代安全带得到普及，许多生命本可挽回。然而，汽车行业并未主动承担责任，而是由国会在当年秋天出手：先设立交通部，随后通过《国家交通与机动车安全法》和《高速公路安全法》，赋予联邦政府制定安全标准的广泛新权力。到1968年，所有新车必须配备安全带。此后60年间，尽管公路交通量大幅增长，道路死亡人数却显著下降。

这段历史让人联想到当下的人工智能安全监管。本周二，多位AI企业CEO齐聚美国总统特朗普身边，互相祝贺签署了一份AI安全协议。我的同事Maddy Varner详细报道了协议的具体条款，但简而言之，这份协议很大程度上依赖AI实验室自我约束。特朗普周三在椭圆形办公室说得更直白：“他们会监管，他们会互相监管，他们会自我监管，他们会管好自己。一切都会非常顺利。”

这就是当前AI安全监管的现实。一边是AI高管、员工和科技界人士，他们都在呼吁需要清晰、审慎、有针对性的规则来避免潜在灾难。另一边是特朗普的一厢情愿——仿佛只要反复说某件事，它就会成真。

这并不是说问题简单。如何恰当地“控制前沿”仍是未解难题。即便所有主要AI实验室同意放缓，也没有有效手段确保他们守信，而中国很可能继续走自己的路。美国经济越来越依赖AI产业支撑，该产业本身已沦为融资交易和循环投资的“衔尾蛇”；任何有意义的打压都可能引发远超硅谷的严重不稳定。

汽车行业的类比也有诸多不完美之处。AI没有安全带那样的对应物，没有能一夜之间显著降低风险的直接方案。潜在危害更不确定、更难以量化。AI末日论者会说这项技术可能十年内导致人类灭绝；怀疑者则认为那是科幻式的营销噱头。

没人能确定接下来会发生什么。如果你在寻找计算个人“末日概率”的数据点，不妨看看CNN最近的报道：美军据称差点登上中国船只，部分原因是一份AI生成的虚假报告称该船载有核材料。这份AI协议在防止类似事件上能起什么作用？不清楚。

面对所有这些不确定性，正确的答案显然不是“走着瞧”。但这正是自我监管所鼓励的。平心而论，OpenAI出于安全考虑推迟甚至取消了最先进模型的发布。但在此之前，这些模型已进行了各种大规模黑客攻击，新细节仍在不断流出。

在《时代》周刊周四发表的采访中，特朗普暗示司法部和联邦调查局足以充当AI行业的“护栏”。真的够吗？当AI模型失控时，AI公司能否被追究法律责任尚不明确。AI代理本身更不会收到传票。

事实是，目前没人知道如何解决AI安全问题。这需要大量协调、资源和监管权力来制定并执行正确框架，而这正是政府理应提供的。特朗普政府却选择袖手旁观，告诉大家踩油门。不用系安全带了。

中文翻译：

给你上一堂有趣的历史课。别担心，我们会讲得很快！1966年，近5.1万人死在美国公路上。大量证据表明，如果配备现代安全带，许多生命本可以挽救，而安全带当时已经问世好几年了。汽车行业的回应是团结起来，以自我监管的胜利姿态，自愿将这一安全装置安装到每一辆下线的新车上。

开玩笑的！他们当然没那么做。1966年秋，国会成立了交通部，随后通过了《国家交通与机动车安全法》以及《公路安全法》，赋予联邦政府制定安全标准的广泛新权力。到1968年，每辆新车都被要求配备安全带。尽管过去60年公路交通量大幅增长，如今道路上的死亡人数却少得多。

我在看多位AI首席执行官周二围坐在美国总统唐纳德·特朗普身边，互相祝贺签署了一份AI安全协议时，想到了这段历史。我的同事玛迪·瓦纳在这里写了具体的条款，但简而言之，很大程度上要靠AI实验室自我约束。

或者，正如特朗普周三在椭圆形办公室更简洁地说的那样：“他们会监管，他们会互相监管，他们会自我监管，他们会管好自己。一切都会非常顺利。”

这就是AI安全监管的现状。一边是AI高管、员工和科技名流，他们都在说需要引入清晰、深思熟虑、有针对性的规则来避免潜在灾难。（比如比尔·盖茨一直在媒体上四处奔走，像卡珊德拉一样预言AI末日。）另一边是特朗普的魔法思维——只要一件事你说得够多，它就一定是真的。

这并不是说这一切简单或容易。如何恰当地“把控前沿”，正如那些高管中一些人喜欢说的那样，仍然没有答案。即使所有主要AI实验室都同意放慢速度，也没有好办法让他们信守承诺，而中国很可能无论如何都会走自己的路。美国经济越来越依赖AI产业支撑，而AI产业本身已经变成融资交易和循环投资的衔尾蛇；任何有意义的打击都可能引发远超硅谷的重大不稳定。

诚然，汽车行业也不是完美的类比，原因有很多。AI没有安全带的对等物，没有一种直接了当的解决方案能够一夜之间明显降低风险。（大多数州也花了几十年才通过法律强制人们真正系安全带。）潜在危害更不确定，也更难以量化。AI末日论者会告诉你，这项技术可能十年内导致人类灭绝；怀疑者则认为那是科幻营销的胡说八道。

我们之前讨论过各种AI风险情景，但没人能确定接下来会发生什么。如果你在寻找数据点来计算你自己的个人“末日概率”，我会指向CNN最近的一篇报道，内容是关于美军据称差点登上一艘中国船只，部分原因是一份AI生成的虚假报告称该船载有核材料。那份AI协议做了什么来防止类似事件在未来发生？不清楚。

面对所有这些不确定性，正确的答案肯定不是“那就走着瞧吧”。但这正是自我监管所鼓励的。值得肯定的是，OpenAI出于安全担忧推迟甚至取消了其最先进模型的发布。但在此之前，它们已经进行了各种大规模黑客攻击，关于这些的新细节仍在源源不断地流出。

在《时代》周刊周四发表的一篇采访中，特朗普表示，司法部和联邦调查局足以充当人工智能行业的“护栏”。是吗？目前远不清楚，当AI公司的模型失控时，它们是否甚至能被追究法律责任。而且AI代理本身又不会被送达传票。

事实是，现在没人知道如何解决AI安全问题。这需要大量协调、资源和监管权力来制定正确的框架并加以执行，而这正是政府被设计来提供的东西。特朗普政府甚至没有尝试去推进任何这类工作，而是选择让路，并告诉所有人踩油门。不用费心系安全带了。

这是《模型行为》通讯的一期。在此阅读往期通讯。
评论
返回顶部

英文来源：

Here’s a fun history lesson for you. Don’t worry, we’ll make it quick! In 1966, nearly 51,000 people died on US highways. The overwhelming evidence suggested that many lives could have been saved with modern seat belts, which by then had been around for several years. The automotive industry responded by rallying together, and, in a triumph of self-regulation, voluntarily put the safety feature in every new vehicle that rolled off the line.
Kidding! Of course they didn’t do that. In the fall of 1966, Congress established the Department of Transportation and then passed the National Traffic and Motor Vehicle Safety Act as well as the Highway Safety Act, giving the federal government sweeping new authority to set safety standards. By 1968, seat belts were required in every new car. Despite a massive increase in highway traffic in the last 60 years, far fewer fatalities happen on the road today.
I was thinking about this while watching various AI CEOs huddle around US president Donald Trump on Tuesday, congratulating themselves and each other on having signed an AI safety accord. My colleague Maddy Varner wrote about the specific stipulations here, but the short version is that it’s largely up to the AI labs to keep themselves in check.
Or, as Trump put it even more succinctly in the Oval Office on Wednesday: “They’re gonna police, and they’re gonna police each other, and they’re gonna self-police, they’re gonna police themselves. It’s going to work out very well.”
So this is the current state of AI safety regulation. On one side are the AI executives, employees, and tech luminaries, who are all saying that there need to be clear, thoughtful, tactical rules introduced to avoid potential disaster. (Bill Gates, for instance, has been on a media blitz playing Cassandra about an AI apocalypse.) On the other side is Trump’s magical thinking that if you say something enough times, it must be true.
That’s not to say any of this is simple or easy. The question of how to properly “pace the frontier,” as some of those same executives like to say, remains open. Even if all of the major AI labs were to agree to a slowdown, there’s no good way to keep them to their word, and China would likely continue on its own path regardless. The US economy is increasingly propped up on the AI industry, which itself has become an ouroboros of financing deals and circular investments; any meaningful crackdown risks major instability that reaches well beyond Silicon Valley.
The automotive industry is also admittedly an imperfect analogy for all kinds of reasons. There’s no seat belt equivalent for AI, no straightforward solution that can demonstrably reduce risk overnight. (It also took decades for the majority of states to pass laws mandating that people actually wear them.) The potential harms are less certain and less quantifiable. AI doomers will tell you the technology could result in human extinction in a decade; skeptics argue that’s sci-fi marketing malarkey.
We’ve talked about various AI risk scenarios before, but no one can say for certain what’s going to happen next. If you’re looking for data points to calculate your own personal p(doom), I’d point to a recent CNN report about how the US military supposedly nearly boarded a Chinese ship in part because a false AI-generated report said it was carrying nuclear materials. What, if anything, does the AI accord accomplish to prevent similar incidents from happening in the future? Unclear.
The correct answer to all this uncertainty surely isn’t “I guess we’ll find out.” But that’s exactly what self-regulation encourages. To its credit, OpenAI has delayed and even canceled the release of its most advanced models over security concerns. But not before they went on various, extensive hacking sprees, new details about which are still trickling out at a steady pace.
In an interview with TIME published on Thursday, Trump suggested that the Department of Justice and the Federal Bureau of Investigation were sufficient “guardrails” for the artificial intelligence industry. Are they? It’s far from clear if AI companies can even be held legally responsible when their models go rogue. And it’s not like AI agents themselves are being served subpoenas.
The truth is, no one knows how to solve for AI safety right now. It’s going to take massive amounts of coordination, resources, and regulatory authority to develop the right framework and enforce it, which is exactly the kind of thing governments are designed to provide. Instead of even attempting to work through any of that, the Trump administration is getting out of the way and telling everyone to step on the gas. Don’t bother buckling up.
This is an edition of the Model Behavior newsletter. Read previous newsletters here.
Comments
Back to top