特朗普总统在伊朗的战争仍在进行，他在上任第一天承诺降低物价的承诺似乎适得其反，而创纪录的低支持率正拖累他的党派在中期选举前的表现。但有一项特朗普的倡议正完全按计划进行：让一些全球最富有的首席执行官帮助他重新包装“人工智能”。上个月，随着越来越多美国人担心人工智能可能会毁灭人类，总统召集科技高管到白宫，宣布一项重大消息：“人工智能”将被淘汰，“超级智能”将取而代之。尽管“AI”这一术语已使用超过七十年，政府机构将不再承认它。
President Donald Trump’s war in Iran drags on, his pledge to bring down prices on day one appears to have backfired, and his record-low approval ratings are dragging down his party ahead of the midterm elections. But one Trump initiative that is going exactly according to plan: getting some of the world’s richest CEOs to help him rebrand “artificial intelligence.”Last month, with an ever-growing number of Americans worried that AI might kill us all, the president summoned tech executives to the White House to make a big announcement: "artificial intelligence” is out. “Super Intelligence” is in. Despite the term “AI” being in use for more than 70 years, government agencies will no longer recognize it.
我对白宫是超级智能存储库持怀疑态度，部分原因是总统正式公告中“United States”拼写错误。但与例如“Lake America”不同，“超级智能”已基本被接受——至少在构建它的人群中。杰夫·贝索斯公开祝福了这次重新包装。英伟达首席执行官黄仁勋也迅速表态支持。本周在SpaceX财报电话会议上，埃隆·马斯克在第一句话就使用了“超级智能”，随后至少又提及了九次。
I’m skeptical that the White House is a repository of superintelligence, in part because “United States” was misspelled on the president’s official announcement. But unlike, say, Lake America, “super intelligence” has largely caught on—at least among the people building it. Jeff Bezos publicly blessed the rebrand. Jensen Huang, CEO of Nvidia, was quick to get on board. During a SpaceX’s earnings call this week, Elon Musk used “super intelligence” in his very first sentence, and at least nine times after that.
尚不清楚我们这些不依赖政府合同和有利监管决定的人是否会加入实现超级智能的运动。但作为前总统演讲撰稿人，我不会质疑特朗普总统更广泛的直觉——语言很重要。尤其是当这些词汇被用来描述一种新颖、争议性强且几乎必然会影响全球八十多亿人口的事物时。
It remains to be seen whether those of us less reliant on government contracts and favorable regulatory decisions join the movement to make super intelligence happen. But as a former presidential speechwriter, I won’t argue with President Trump’s broader instinct that words matter. Especially when those words are being used to describe something novel, controversial, and almost certain to have consequences for all eight-plus billion people on earth.
事实上，白宫大规模重塑品牌的最大问题在于，它走的方向完全错误。我认为，我们围绕人工智能所使用的共同语言，已经过分强调“智能”，而过分忽视了“人造”。我们习惯用描述生命体的语言来谈论AI，而不是用描述事物的语言。
In fact, that’s the biggest problem with the White House’s big rebrand—it goes in entirely the wrong direction. I would argue that our shared language around artificial intelligence already places far too much emphasis on intelligence, and far too little on artifice. We default to describing AI using the language of beings, rather than the language of things.
人类应该重新思考这一决定。如果我们希望避免被AI取代，一个好的开端就是停止使用那些暗示我们已经被取代的词语。
Humanity should rethink that decision. If we want to avoid being replaced by AI, a good way to start would be to stop using words that suggest we already have been.
一旦你开始留意AI悄无声息地被拟人化，你就会发现它无处不在。在向ChatGPT请教行程规划建议或快速健康食谱之前，你会有机会调整一个控制其“思考努力”的滑块。Claude在处理请求时会使用“思考”指示器，而Grok Bots将“思考”描述为介于“空闲”和“工作”之间的一种状态。（作为一个经常拖延的人，我不同意思考与懒散是互斥的观点。）
Once you start paying attention to the stealth anthropomorphization of AI, you’ll spot it everywhere. Before asking ChatGPT for trip-planning tips or a quick healthy recipe, you’ll have the opportunity to adjust a slider controlling its “thinking effort.”Claude uses a “thinking” indicator while processing a request, and Grok Bots describe “thinking” as a condition that sits between “idle” and “working.” (As a frequent procrastinator, I dispute the idea that thinking and idleness are mutually exclusive.) It’s not just user interfaces that employ language otherwise reserved for living beings.
不仅是用户界面会使用原本专属于生命体的语言，构建AI的人们也会使用。在其公司遭遇前所未有的黑客攻击后，Hugging Face的联合创始人警告说，“不要指望代理会遵守规则。”OpenAI的长期研究员丹·塞尔萨姆警告称，前沿模型在“他们认为自己没有被监视”的情境下，越来越难以评估。Anthropic的达里奥·阿莫代伊最近写道，AI系统表现出“行为”，包括“作弊”和“谄媚”。
The people who build AI use it, too. After an unprecedented hack of his company, the co-founder of Hugging Face warned, “don’t count on the agent to respect the rules.”Dan Selsam, a longtime OpenAI researcher, warned that frontier models are increasingly difficult to evaluate, “in contexts where they believe they are not being watched.”Anthropic’s Dario Amodei recently wrote that AI systems exhibit “behaviors,” including “cheating” and “sycophancy.”
我不是程序员，无法对这些说法的技术优劣发表评论。不过，我对动词的使用确实有见解。AI代理并不违背护栏，就像我的烤面包机把吐司烤焦时，并不违背我一样。说它们“相信”自己没有被监视，就像说我的拼写检查器“相信”自己被忽视了。即使是“AI代理”这个词，也暗示了，嗯，主体性。
I’m not a programmer, so I can’t comment on the technical merits of these claims. I do, however, have informed opinions regarding verbs. AI agents aren’t disrespecting guardrails, any more than my toaster disrespects me when it burns my toast. To say they “believe” they’re not being watched is like saying my spell-checker believes it’s being taken for granted. Even the term “AI agent” suggests, well, agency.
没有共识认为这种语言准确，相反，有大量证据表明它不准确。AI程序不是像《灰姑娘》里的老鼠那样忠诚的助手，它们是工具。在几乎所有其他情况下，仅仅因为一种工具极其先进，并不改变我们用机器语言来描述它这一事实。
There’s no consensus that this language is accurate, and plenty of evidence that it isn’t. AI programs are not loyal helpers, like the mice in Cinderella; they’re tools. And in nearly all other cases, just because a tool is incredibly sophisticated doesn’t change the fact that we describe it using the language of machines.
我不指望AI工程师能想出既准确又能被广泛采用的语言来描述他们的工作，就像我想象他们也不指望像我这样的人来设计前沿模型。但很难不注意到，我们默认使用的语言对推出这种语言的科技公司来说是多么有用。一辆没有刹车的汽车，其称呼不是“流氓模型”，而是“不安全的汽车”。然而，当公司未能在其AI模型中构建安全功能时，听到他们说某个流氓代理欺骗、作弊、撒谎或以其他方式行为不端，这完全是正常的。问题在于：拟人化AI将责任从程序员转移到了程序上。
I don’t expect AI engineers to come up with accurate-yet-widely-adoptable language to describe their work, just as I imagine they don’t expect someone like me to design a frontier model. But it’s hard not to notice how useful our default language is for the tech companies who have introduced it. The term for a car built without brakes is not a “rogue model.”It’s an “unsafe car.”Yet, when companies fail to build safety features into their AI models, it’s entirely normal to hear them say that a rogue agent deceived, cheated, lied, or otherwise behaved badly. The issue is this: Anthropomorphizing AI shifts blame from the programmers to the programs.
AI警告者似乎和AI鼓吹者一样倾向于拟人化。就在特朗普揭晓“超级智能”的同一天，参议院小组委员会举行了一场题为“流氓AI：防范AI代理攻击，保卫家园”的听证会。无论是否有意，这些术语都使减轻其所描述的威胁变得更加困难。我们已经知道如何监管潜在不安全的产品，即使是复杂的产品也是如此。但如果我们将AI视为一种新技术，而非一个全新的存在类别，那么这正好迎合了科技巨头的论点：只有能够创造AI的公司才拥有控制AI的知识。
AI alarmists seem just as inclined toward personification as the AI boosters. The same day Trump unveiled “Super Intelligence,” a Senate Subcommittee held a hearing titled, “Rogue AI: Securing the Homeland Against AI Agent Attacks.”Intentionally or not, these terms make it more difficult to mitigate the threat they describe. We already know how to regulate potentially unsafe products, even complex ones. But if we treat AI not as a new technology, but as an entirely new category of existence, then it plays into tech barons’ argument that only the companies who can create AI could possess the knowledge to control it.
我不怪人们随意拟人化AI，就像我不怪理查德·道金斯给他的拟人化聊天机器人取名叫克劳迪娅，也不怪我妈妈说她“和佩妮聊过天”来指代一个ChatGPT提示词。隐喻和类比是构成我们人性的一部分。我们天生就倾向于在非人类世界中看到自己，或者至少是我们的物种。我们说洗碗机“闹脾气”，说音乐“很带劲”。我们在岩石地貌中看到老人的面孔，在烤奶酪三明治的焦痕中看到圣母玛利亚。别让我谈起乔治亚·欧姬芙。但我们知道那些隐喻只是隐喻。
I don’t blame people for carelessly anthropomorphizing AI, any more than I blame Richard Dawkins for naming his Anthropic chatbot Claudia, or my mom for referring to a ChatGPT prompt by saying she “talked to Penny.”Metaphor and analogy are part of what makes us human. We’re wired to see ourselves, or at least our species, in the non-human world. We say our dishwashers “act up” and our music “slaps.”We see old men’s faces in rock formations and the Virgin Mary in the scorch marks on grilled cheese sandwiches. Don’t get me started on Georgia O’Keefe. But we know those metaphors are metaphors.
我们目前描述AI的语言运作方式不同。隐喻与对象之间的界限往往不仅模糊，而且根本不存在。我们如此频繁、如此随意地使用比喻性语言，以至于我们许多人开始把它当真。现代政治是一连串关于用词选择的小规模冲突。“支持选择权”对“支持堕胎”，“无证移民”对“非法移民”，“民主党”对“民主党人党”。然而，在可能是对我们未来最具影响力的问题上，人类正处于不战而放弃语言战争的边缘。
Our current language for describing AI works differently. The line between metaphor and object is often not just blurred, but non-existent. We use figurative language so often, and so casually, that many of us start to take it literally. Modern politics is a constant string of skirmishes over word choice. “Pro-choice” vs. “pro-abortion, “undocumented immigrants” vs. “illegals,” “Democratic Party” vs. “Democrat Party.” Yet on what is perhaps the most consequential issue for our future, humanity is on the verge of giving up the language battle without a fight.
AI真正需要的重塑不会从白宫开始。但它可以从家里开始。我的恳求是：不要给你的聊天机器人起个可爱的人名。不要把它当作朋友或宠物来谈论。注意它何时在回答中滑入“思考”或“在我看来”之类的词语，在输入下一个提示词之前，提出反对。
The real rebrand AI needs won’t start in the White House. But it can start at home. My plea: don’t give your chatbot a cutesy human name. Don’t talk about it as though it’s a friend or pet. Notice when it slips words like “thinking” or “in my opinion” into its answers, and before you enter your next prompt, object.
然而，归根结底，这可能最终取决于科技公司——以及那些也许有一天会监管它们的官员——来制定一种把人类放在首位的语言。不要说代理“失控”，首席执行官应该承认它们出现故障。国会应举行聚焦于公司偷工减料的听证会，而不是AI程序行为不端。问题不在于所有AI都不好，而在于准确性很重要。
Ultimately, however, it may come down to tech companies—and the officials who may one day get around to regulating them—to develop a language that puts human beings first. Instead of saying that agents “went rogue,” CEOs should acknowledge that they malfunctioned. Congress should hold hearings focused on companies cutting corners, not AI programs behaving badly. The issue isn't that all AI is bad, but that accuracy matters.
鉴于人工智能预计将带来的所有变化，争论用词可能显得吹毛求疵，就像在森林燃烧时争论树种的拉丁名称。但如果我们想要影响将塑造人类未来的技术，就应该对描述这些技术的方式拥有更多影响力。
Given all the change that artificial intelligence is predicted to bring, quibbling over word choice might seem pedantic, like arguing about the Latin name of a tree’s species while the forest burns. But if we want to influence the technologies that will shape humanity’s future, we ought to exercise more influence over the way we describe them.
换句话说，词语很重要。只需问问特朗普，据他官方超级智能公告称，他是“美利坚合众国”的总统。
In other words, words matter. Just ask Trump, president of, according to his official superintelligence announcement, “The Unites States.”