聊天机器人正将我们推向后人类互联网

Photo illustration by Chantal Jahchan

Remember job applications? You know, you’d write a résumé and a cover letter. Someone would read them. You’d fish out a tasteful button-down and try to charm your interviewer. Maybe your one-liner or a question about your amateur bowling career would help you score the position. Maybe you’d never hear back.

还记得求职申请吗?你懂的,就是你要写一份简历和一封求职信。然后会有人来读它们。你要翻出一件有品位的纽扣衬衫,努力给面试官留下好印象。也许一句俏皮话或者一个关于自己业余保龄球生涯的问题能帮你拿下这个职位。也许你从此再也得不到回音。

That’s all changing. When Christian Vinson applied for jobs in finance this year, he knew he was really auditioning for artificial intelligence. Many companies now use screening software to sift through piles of applications, so he spent close to 10 hours teaching two A.I. chatbots about his work experience. He filled in details about his education, a banking internship and five years in the Army. He then prompted the bots to tailor his cover letter to hundreds of job listings.

这一切都在改变。今年,当克里斯蒂安·文森申请金融行业的工作时,他清楚自己实际上是在向人工智能求职。现在很多公司都使用筛选软件来筛查堆积如山的求职申请,因此他花了将近10个小时,让两个人工智能聊天机器人了解他的工作经验。他详细填写了自己的教育背景、一段银行业实习经历以及五年军旅生涯的信息。然后,他向机器人发出指令,让它们针对数百个职位招聘量身定制他的求职信。

A.I. had written his applications, and A.I. was reading them too — a dog chasing its own tail. “This is the new normal,” says Vinson, 30, who landed a job at an investment bank in New York.

人工智能写了他的求职申请,而阅读这些申请的同样也是人工智能——就像狗追自己的尾巴一样。“这就是新常态,”30岁的文森说,他最终在纽约一家投资银行找到了一份工作。

We live in a world in which A.I. tools have found their way to the center of a staggering number of professional, academic and personal interactions. Lately these chatbots and agents have been doing more than just talking to us. Students are writing essays with ChatGPT, and teachers are using A.I. to grade them. Workers are sending lengthy A.I.-generated emails to colleagues, who are responding with walls of chatbot text of their own. A.I podcast hosts are holding forth with A.I. podcast guests on topics like gas prices, the Iran war — and artificial intelligence. Tech wants to take all this even further: In March, Meta acquired a social network designed for bots to talk to one another.

我们生活在一个这样的世界里:人工智能工具已经渗透到了数量惊人的职业、学术和个人互动的核心。最近,这些聊天机器人和智能代理所做的已经不仅仅是与我们交谈。学生们用ChatGPT写论文,老师们则用人工智能来给论文打分。员工向同事发送由人工智能生成的冗长电子邮件,同事们则用他们自己生成的长篇聊天机器人文本回复。人工智能播客主持人和人工智能播客嘉宾就汽油价格、伊朗战争以及人工智能等话题滔滔不绝。科技界还想把这一切推得更远:今年3月,Meta收购了一个专为机器人相互交谈而设计的社交网络。

The dead internet theory, a concept popular in tech circles, holds that most digital traffic will eventually be software, leaving humans as a minority voice. We may now have a corollary of that prediction — let’s call it a bot loop — when the humans on both sides of an interaction hand over their role to artificial intelligence. These recursive exchanges can be eerie and error-prone, a single mistake bouncing back and forth uncorrected over and over. Stranger still, they flip our conversational dynamic inside out, turning chatbots into the interlocutors and humans into their enablers.

科技圈流行着一个名为“死网理论”的概念,该理论认为,绝大部分数字流量最终都将由软件产生,人类的声音将沦为少数。我们现在或许已经有了这一预测的推论——姑且称之为“机器人循环”——互动双方的人类都将自己的角色交给了人工智能。这种递归式的交流可能令人毛骨悚然且容易出错,一个简单的错误可能会在未经纠正的情况下反反复复地来回传递。更奇怪的是,它们彻底颠覆了我们的对话动态,把聊天机器人变成了对话者,人类变成了它们的协助者。

Depending on whom you ask, bot loops might be banal, dystopian or genuinely useful. No matter how people feel about them, researchers say we’re going to see bot loops much more often as we deputize A.I. systems to do our bidding.

取决于你问的是谁,机器人循环可能是平庸无奇的、反乌托邦的,也可能是真正有用的。无论人们对此作何感想,研究人员表示,随着我们日渐委托人工智能系统来执行我们的命令,我们将越来越频繁地看到机器人循环。

Someday in the not-too-distant future, a pipe will burst and you will need a plumber in a hurry. Maybe you will call 10 plumbers yourself to ask about rates — but perhaps you will turn all of that dialing over to a kind of personal digital assistant that can perform tasks like sending emails, buying plane tickets and writing code. Plumbers, meanwhile, may struggle to keep up with an influx of calls from tireless A.I. assistants, says Jiaxin Pei, an assistant professor at the School of Information at the University of Texas at Austin. Maybe they’ll direct an agent of their own to act as a virtual receptionist. “This is this endless loop we’re creating,” Pei says. How will all this affect the way we interact with one another?

在不太遥远的未来某一天,水管可能会爆裂,你急需一名水管工。也许你会亲自给10名水管工打电话询问价格,但可能你会把所有这些拨号工作交给一种个人数字助理,它能执行发送电子邮件、购买机票和编写代码等任务。与此同时,德克萨斯大学奥斯汀分校信息学院助理教授裴家新(音)表示,水管工们可能会难以应付大量来自不知疲倦的人工智能助手的电话。也许他们会指派自己的智能代理来充当虚拟接待员。“这就是我们正在创造的无尽循环,”裴家新说。这一切将如何影响我们彼此互动的方式?

马特·施利希特正在演示“Moltbook”,这是他为人工智能代理创建的一个互联网论坛。

Looked at one way, human history is one long string of freakouts about the potential for new technology to alter our communication. Socrates worried that writing would wreck our memory — something we know only because Plato wrote it down. Early-20th-century critics argued that the telephone posed a threat to civility. Repeatedly, detractors have predicted that the quality of our interactions might suffer as a result of modern convenience. The author of a 1994 article in The New York Times argued that electronic mail was “so compelling and easy that it threatens to overwhelm its usefulness through sheer volume and lack of consideration,” although he conceded that it did have its perks: “There are no stamps to lick.”

从某种角度来看,人类历史就是一长串对于新技术可能改变我们交流方式的恐慌史。苏格拉底曾担心书写会毁掉我们的记忆——我们之所以知道这件事,只是因为柏拉图把它写了下来。20世纪初的批评家们认为,电话对文明构成了威胁。批评者们一次又一次地预测,现代生活的便利可能会导致人际互动质量下降。《纽约时报》1994年一篇文章的作者认为,电子邮件是“如此引人注目且便捷,以至于其海量信息和缺乏考虑的特性,可能使其有用性大打折扣”,尽管他也承认电子邮件确实有好处:“再也不用舔邮票了。”

None of these innovations brought about the end of productive discussion, although each has changed its texture. We can now trade messages with more people than ever without needing to be in the same physical space. As a result, researchers think we may actually be speaking to one another less. Many of our conversations take place on platforms that incentivize certain kinds of exchanges, particularly those that are heated, outlandish or otherwise likely to retain our attention for the platforms’ financial gain.

这些创新没有任何一项终结了富有成效的讨论,尽管它们各自都改变了讨论的质感。如今,我们无需身处同一物理空间,就能与比以往更多的人交流信息。结果,研究人员认为,我们彼此之间的实际交谈可能变少了。我们的许多对话都发生在一些平台上,这些平台会鼓励某些类型的交流,尤其是那些激烈的、古怪的,或者其他有可能为了平台的经济利益而留住我们注意力的交流。

Still, A.I. is different from these previous advancements. While email transmits messages between people, chatbots emulate people — and perhaps replace them. All of a sudden, humans get to decide whether they would prefer to have a conversation with a person or with a program that approximates one. Plenty of people are deciding that they would prefer to go to a chatbot for financial advice, medical diagnoses or companionship.

尽管如此,人工智能与以往的这些进步依然有所不同。电子邮件是在人与人之间传递信息,聊天机器人则是在模仿人类,甚至可能取代人类。突然之间,人类要开始自行决定,他们是更愿意与真人交谈,还是与一个近似真人的程序交谈。很多人已经做出决定,他们更愿意向聊天机器人寻求财务建议、医疗诊断或陪伴。

People are already using A.I.-generated avatars to speak for them. When Justin Lester, a pastor in the Bay Area, isn’t available, people can speak with his “digital twin.” The next step is very likely a future in which the people’s own digital twins might speak with Lester’s. Increasingly, our digital intermediaries will collide. We’ll talk to our bots, and our bots will talk to one another. Chief executives have created A.I.-powered avatars that can meet with their employees. How long will it be before their underlings send their avatars to those meetings, too?

人们已经在使用人工智能生成的虚拟化身来替他们发言。当湾区牧师贾斯汀·莱斯特无暇分身时,人们可以与他的“数字孪生体”交谈。下一步很可能是这样的未来:人们自己的数字孪生体可能会与莱斯特的数字孪生体交谈。我们的数字中间人将越来越多地发生碰撞。我们会与我们自己的聊天机器人交谈,我们的聊天机器人也会相互交谈。首席执行官们已经创建了由人工智能驱动的虚拟化身,这些化身可以同他们的员工开会。距离他们的下属也派虚拟化身去参加这些会议,还要多久呢?

加利福尼亚州瓦列霍市友谊浸信会的贾斯汀·莱斯特牧师训练了一个人工智能代理,以便在他无法接见教友时提供咨询。

Plenty of people are already living in this strange new world. I’ve spent the past couple of months interviewing them. There was a graduate student completing work with A.I. that was being graded by his professor’s virtual assistants. There was a telecom executive who had noticed customers sending A.I. voice agents to argue with his company’s A.I. customer service system. There was a man who used a chatbot to draft messages on a dating app to a woman who seemed to be doing the exact same thing. “How long are we gonna ChatGPT each other?” he eventually wrote. She never responded.

许多人已经生活在这个奇妙的新世界里。在过去的几个月里,我采访了他们。有一名研究生用人工智能完成作业,而给他打分的是教授的虚拟助手。有一位电信业高管注意到,有客户派出人工智能语音代理与他公司的人工智能客服系统争吵。还有一名男子在约会软件上使用聊天机器人给一名女子起草消息,而那名女子似乎也在做完全相同的事情。“我们要互相ChatGPT多久?”他最终写道。但她再也没有回复。

Most of these people acknowledged a certain hollowness to the interactions, which followed the basic contours of a human conversation but got the texture all wrong. A single bot creates mountains of text that are dazzling in their phoniness, like costume jewelry. Stick two of them together in conversation and they generate an endless flow of dialogue that follows all the rules yet still stands out for what it is missing: curiosity, empathy, actual experience.

其中大多数人都承认,这种互动存在一种空洞感,它们虽然遵循了人类对话的基本框架,但质感却完全不对。一个单独的聊天机器人能生成海量文本,其虚假程度令人眼花缭乱,宛如廉价首饰。把两个机器人放在一起对话,它们会产生源源不断的对话流,虽然遵循所有规则,却因缺少某些关键要素而显得突兀:好奇心、同理心和真实的经验。

The risks of these loops are still coming into view, but one big problem is just how frequently chatbots get things wrong. When two A.I. systems interact, particularly ones powered by the same large language models, they may miss each other’s errors and eventually amplify them, says Soheil Feizi, an associate professor of computer science at the University of Maryland. “They are just sharing the same blind spots,” he says.

这种循环的风险仍在显现,但一个大问题是,聊天机器人犯错的频率实在是太高了。马里兰大学计算机科学副教授苏海尔·菲齐表示,当两个人工智能系统交互时——尤其是由相同大型语言模型驱动的系统,它们可能会忽略彼此的错误,并最终导致这些错误被放大。“它们只是在共享相同的盲点,”他说。

A recent paper by a researcher at Harvard Medical School considered how this might play out in a hospital setting. Say one A.I. tool is responsible for analyzing X-rays to label broken bones. It sends an assessment to another A.I. tool that assigns rooms and a third one that determines the order of patient treatment. If an initial scan is incorrect and never subject to human verification, the error could echo through the entire network and affect the patient, the doctor and her clinic.

哈佛大学医学院一位研究人员最近发表了一篇论文,探讨这种情况在医院环境中可能会如何演变。假设一个人工智能工具负责分析X光片以标记骨折情况。它将评估结果发送给另一个负责分配病房的人工智能工具,以及第三个决定患者治疗顺序的工具。如果最初的扫描结果有误,且从未经过人工核实,这个错误就会在整个网络中产生连锁反应,影响到患者、医生及诊所。

An even sneakier issue is that artificial intelligence systems seem to be biased toward their own work. One study found that A.I. résumé screeners preferred applications written by A.I. to ones written by humans. (In that way, Vinson, the chatbot-happy job applicant, may have helped his own chances.) This phenomenon may already be wreaking havoc on the college admissions process, in which some schools use A.I.-powered essay readers to score submissions that, no doubt, are often generated by chatbots.

一个更隐蔽的问题是,人工智能系统似乎对人工智能生成的内容存在偏袒。一项研究发现,人工智能简历筛选器更喜欢人工智能撰写的申请,而不是人类撰写的申请。(从这个意义上说,喜欢使用聊天机器人的求职者文森可能已经提升了自己成功的几率。)这种现象可能已经对大学招生过程造成了严重破坏——一些学校使用人工智能驱动的文章阅读器来给申请材料打分,毫无疑问,这些材料往往也是由聊天机器人生成的。

Despite these shortcomings, Sarah Davis, a cultural anthropologist and the dean of St. John’s College in Santa Fe, N.M., wonders whether people might grow so accustomed to the smoothness of A.I.-to-A.I. interactions that they could lose their tolerance for the necessary unpredictability of human relationships. She says she worries about chatbots allowing a “kind of ease that lets us become worse versions of ourselves, and a kind of alienation, a lack of togetherness that requires that we face friction.”

除了这些缺陷,新墨西哥州圣达菲市圣约翰学院的文化人类学家兼院长莎拉·戴维斯还想知道,人们是否会习惯于人工智能之间互动的顺畅,因此失去对人际关系中必要的不可预测性的包容。她说,她担心聊天机器人会带来一种“让我们沦为更差版本的自己的安逸感,以及一种异化,不再有那种需要我们直面摩擦的团结感”。

It might be tempting to think of the bot loop as a stopover on the way to some totally autonomous world. Perhaps A.I.-generated “counterfeit people,” as the philosopher Daniel C. Dennett called them, will work together to create even more counterfeit people, exploiting our dependence on A.I. systems and putting human civilization itself at risk. The likelier outcome is that the share of interactions that are between humans will just shrink as the share of bot-to-bot interactions grows — a profound dilution that affects the way we move through the world.

人们或许会倾向于认为,机器人循环只是通往某个完全自主世界途中的一个中转站。或许人工智能生成的“伪人”(如哲学家丹尼尔·C·丹尼特所称)将协同合作,创造出更多“伪人”,利用我们对人工智能系统的依赖,将人类文明本身置于危险之中。但更有可能的结果是,随着机器人之间互动比例的增加,人类之间互动的比例将会单纯地萎缩。这是一种深刻的稀释,会影响我们在世界上前行的方式。

In that case, the human voice on the phone that once was a reasonable expectation might become a rare privilege, something akin to a luxury good. Sure, you can talk to a flesh-and-blood receptionist, but it will cost extra. Paradoxically, the people most likely to be able to afford such a luxury might be the ones being enriched by the A.I. boom. Will they use their spoils to get access to the kinds of conversations they are making rarer for everybody else? That might just be the most insidious loop of all.

在那种情况下,电话那头曾经被视为理所当然的人类声音,可能会变成一种罕见的特权,类似于一种奢侈品。当然,你可以同有血有肉的人类接待员交谈,但这需要额外付费。矛盾的是,最有可能负担得起这种奢侈的人可能正是那些因人工智能热潮而致富的人。他们会用自己的战利品去获取那些他们令普通人越来越难以企及的对话吗?这可能才是所有循环中最阴险的一个。

相关报道

最受欢迎
  1. 郭德纲改编红歌被调查,中国言论空间日益收窄

  2. 恒大创始人许家印被判无期徒刑

  3. 中国首次实现火箭陆地回收

  4. 如果北京武力攻台,普通人能“保卫台湾”吗?

  5. 前福奇顾问承认试图隐瞒与新冠起源相关的记录

  6. 经济悲观情绪蔓延,中国农民工文学引发共鸣

  7. 朱镕基葬礼在北京举行,当局谨慎应对公众情绪

  8. 美国债务突破40万亿美元,借贷狂潮继续

  9. 欧洲加大对前沿技术投资,能否与美中抗衡?

  10. 攻击朋友、讨好敌人:特朗普“黑白颠倒”的世界观

免费下载 纽约时报中文网
iOS 和 Android App

点击下载iOS App 点击下载Android App© 2026 The New York Times Company.

内容来自 纽约时报双语版,版权归原作者所有。

阅读原文 ↗

把喜欢的内容,留在你的新标签页。

打开我的订阅