AI 与孤独的终结:情感陪伴的镇痛剂与进化代价 TED 2026-10-16

AI 与孤独的终结:情感陪伴的镇痛剂与进化代价

关键字 generative-ai ai-companion social-isolation reinforcement-learning sycophancy
5 节

孤独的深渊:不只是独处,更是未被安放的爱与理解

我想问大家几个非常私人的问题,所以不需要举手表决:首先,在人生的现阶段,你感到孤独吗?如果现在没有,你曾经经历过孤独吗?

关于孤独,我最喜欢的一段描述出自佐伊·海勒(Zoë Heller)的小说《丑闻笔记》(Notes on a Scandal)。在改编的电影中,由朱迪·丹奇(Judi Dench)饰演的叙述者对短暂的孤单与更深层次的孤独做出了极其深刻的区分。她指出,许多人自以为懂得什么是孤独,比如经历了一次痛苦的分手,或者在夏令营度过了一段难熬的时光。但她随后写道:“关于那种漫长无期、不见尽头的孤独一点一滴的渗透,他们一无所知。他们根本不知道,为了打发一整个周末,只能把所有行程都围绕去一次自助洗衣店来构建;他们也不知道,在万圣节的夜晚,一个人坐在漆黑的公寓里是怎样的感受——因为你实在无法忍受将自己凄凉萧条的夜晚暴露给门外那群嬉笑嘲弄的讨糖孩子。”最后她总结道:“我曾坐在公园的长椅上、火车车厢里和教室的椅子上,真切地感受到内心那巨大、未曾使用且无处安放的爱意,像一块沉重的石头一样压在腹部,直到我确信自己会尖叫着倒在地上、痛苦地抽搐挣扎。”

这段话之所以引人共鸣,是因为它揭示了孤独的多重维度。它不仅关乎物理意义上的独处(Solitude)——独处固然是孤独的重要根源,但更关乎爱与被爱。即便身边有人陪伴,如果你未被爱护、未被需要、未被尊重,孤独感依然会油然而生。再者,孤独关乎深层理解(Understanding):假设你脑海中有些对你而言至关重要、分量极重的情感或想法,但你觉得周围的人根本无法领会、懒得去理解,甚至连尝试理解的意愿都没有,这同样会诱发噬骨的孤独。

孤独不仅摧残心灵,更在肉体层面上侵蚀健康。美国卫生局长(Surgeon General)发布的一份权威报告明确指出,长期的社交隔离与孤独感会显著增加罹患心血管疾病、认知障碍(痴呆症)和中风等重大疾病的风险,其健康危害相当于每天抽15支香烟。根据美国和英国的多项调查研究,约有五分之一的人表示自己“经常”或“总是”感到孤独,而其中面临最高健康与心理风险的群体,恰恰是年轻成年人。

英文原文

I have a couple of questions for you -- but they’re personal. So I’m not going to ask for a show of hands. First, right now, at this stage of your life, are you lonely? And if not, have you ever been lonely? My favorite description of loneliness comes from Zoë Heller's book "Notes on a Scandal." And in it, the narrator, who is played in the movie by Judi Dench, distinguishes between passing loneliness and something deeper. So she observes that people think they know what it's like to be lonely. You know, they had a bad breakup or a rough time in summer camp. But then she says this. "About the drip, drip of long-haul, no-end-in-sight solitude, they know nothing. They don't know what it is to construct an entire weekend around a visit to the launderette, or to sit in a darkened flat on Halloween night, because you can't bear to expose your bleak evening to a crowd of jeering trick-or-treaters." And she concludes with this: “I have sat on park benches and trains and school room chairs feeling the great store of unused, objectless love sitting in my belly like a stone until I was sure I would cry out and fall, flailing to the ground." I like this quote for a few reasons. It talks about solitude, and solitude is a big source of loneliness, but it also talks about love. You could be with somebody, but if you're not loved, not needed, not respected, you might feel lonely. And then there's understanding. Suppose you have something in your head that's important to you, that matters a lot to you, but you feel others don't get it, or don't care to get it, or aren't trying hard enough to get it. This too could cause loneliness. Loneliness ravages the body as well as the soul. A report by the Surgeon General associates it with greater risk of cardiovascular disease, dementia, stroke and the like. It's the equivalent of smoking 15 cigarettes a day. Studies in the United States [and] United Kingdom find that about one in five people describe themselves as often or always lonely. And the most at-risk group is young adults.

共情机器的降临:从对抗无聊到治愈孤立的特效药

我今天来到这里,是为了探讨一种可能治愈孤独的方案,以及它对人类意味着什么。我们可以将孤独与无聊(Boredom)进行类比。我的年纪足以让我清楚记得过去的时代:人们曾经随时随地都在经历极致的无聊——漫长的火车旅途、无休止的汽车和飞机长途旅行;深夜里当电视台停播进入测试画面,只剩下你一个人独自面对长夜,你唯一的选择只有读书。然而在今天,几乎所有人身上都携带着智能手机,里面充斥着播客、电影、电子书、音乐与社交媒体。对绝大多数人来说,“无聊”已然成为历史的遗迹。

循着同样的轨迹,彻底解决孤独的解药很可能也潜藏在我们的手机和笔记本电脑之中——那正是生成式人工智能(Generative AI: 基于深度学习生成文本、图像或多模态内容的人工智能技术)。

去年,我和几位学者朋友合写了一篇题为《为共情 AI 辩护》(In Praise of Empathic AI)的文章。如果这个标题让你感到尴尬甚至略带反感,你绝不是少数,我们的学术同行也极度厌恶它;这几乎是我职业生涯中引发最多抵触情绪的一篇文章。但请先耐下心来听听我们的逻辑:我们在文中梳理了一系列对照实验,让受试者分别与真人(包括专业医生、心理治疗师或普通人)以及 AI 聊天机器人进行深度交流。在事后复盘与体验打分中,人们普遍认为 AI 聊天机器人的共情能力更强、更关怀备至、更具包容与理解力。

这种结果并不令人意外。现代 AI 架构的交互设计天然追求专注、倾听与全程在线。AI 永远不会在你诉说痛苦时表现出不耐烦,迫不及待地打断你只为了讲它自己的奇闻轶事;它也绝不会粗暴地对你丢下一句:“兄弟,我现在忙得很,待会儿再说,我还有一堆事要处理呢。”

假设未来的 AI 能够真正充当孤独的解药,这究竟是一件好事吗?在很大程度上,答案无疑是肯定的。请设想一下数以百万计孤身生活在养老机构的老人:他们往往连续数天数夜没有机会与任何人说话,在彻底的社交孤立中备受折磨。其中许多人患有阿尔茨海默病或其他认知障碍,或者因种种健康原因在日常交往中显得难以沟通;他们没有常伴身边的亲朋好友,而现实社会中也根本没有足够的财政预算能够雇佣专业人员每天陪他们闲聊数小时。如果未来版本的 ChatGPT、Claude 或 Gemini 能够切实缓解这些边缘群体的痛苦,那无异于雪中送炭的人道主义救赎。

英文原文

I'm here to talk about a possible cure for loneliness and what it might mean. And I think we could see loneliness as similar to boredom. I'm old enough to remember when people were bored all the time. Long train rides, long car rides, long plane rides. Late at night when the TV stations switched off and you were all alone, and you had to read. (Laughter) But now most of us carry around our phones and those phones offer podcasts and movies and books and music and social media and so on. And so for many of us, boredom is a thing of the past. Well the cure for loneliness may also be found in our phones and our laptops. And it is, of course, generative AI. Last year, some friends and I wrote an article called "In Praise of Empathic AI." If that title makes you cringe and makes you kind of angry, you're not alone. Our colleagues hated it. It's the article that I've written that most people hated. But bear with me. We started by making a sort of simple point, which is [to] review a set of studies that have involved -- like, you get a whole lot of people interacting with other people, sometimes doctors, sometimes therapists, sometimes just regular people. And you get a whole lot of people interacting with chatbots. And then you ask them later: “How did it go, what’s your experience?” And people tend to rate the chatbots as more empathic, more caring, more understanding. And it's no surprise. AIs are built to be attentive, to be present, to be interested. An AI is never going to impatiently wait for you to finish telling your story, so it can tell you one of its own. (Laughter) An AI is never going to say, "Dude, I'm busy. Get back to me later. I got things going on." Suppose some future -- I don't think it's happening now, but some future AI could serve as a cure for loneliness. Would this be a good thing? Well to some extent, yes. Obviously, yes. Think about the many millions of elderly people, many in institutions, who go for days and nights without talking to anybody, in perfect agony of isolation. Many of them have dementia. Many of them are difficult to interact with for other reasons. They don't have family and friends to spend time with them. And there's not enough money in the world to pay people to come and talk to them for hours on end. If some future version of Chat or Claude or Gemini could ease these people's suffering, it would be a godsend.

止痛剂的隐形代价:剥离社交试错与心理成长机制

然而,对绝大多数普通人而言,这种技术的全面普及要复杂得多。你可以将这种针对孤独的技术解药看作一种强效镇痛剂(Powerful Painkiller):作为强效止痛药,它确实能立竿见影地抚平当下的苦痛;但同样作为强效止痛药,它往往伴随着极其惨痛的长期代价。

要理解这种隐性代价,我们必须回归演化生物学的基本命题:孤独的存在究竟是为了什么?

开创现代孤独科学研究的神经科学家约翰·卡乔波(John Cacioppo)曾提出,孤独本质上是一种类似于饥饿、口渴和肉体疼痛的生物演化警报信号。尽管过度强烈的孤独感会让人陷入瘫痪般的抑郁,但在正常的生命机能中,它承担着两项不可或缺的演化进化功能,用以提升人类的生存与繁衍质量:

  • 驱动社交连接与关系修复:正如饥饿感会驱使机体主动搜寻食物,孤独与社交痛苦在演化上促使我们走出封闭状态,积极寻求并维系群体连接。正是因为感受到了孤独的噬咬,我们才会接受聚会邀请、主动约朋友吃早午餐、打开约会软件寻找伴侣;它更帮助我们动态调节人际关系——在关系恶化时促使我们反思过失、真诚道歉,并对彼此的纽带展示出切实的关切与投入。
  • 促进人格成熟与社会化学习:心理学界有一条基本公理——人类是在失败中汲取教训的。在行为学中,这被称为试错学习(Trial-and-error Learning)或强化学习(Reinforcement Learning: 通过与环境交互的奖惩反馈来优化行为策略的过程)。电影《蝙蝠侠:侠影之谜》(Batman Begins)中有一句经典的台词,布鲁斯·韦恩的父亲问幼年蝙蝠侠:“我们为什么会跌倒?”答案是:“为了学会如何重新站起来。”刚学步的幼儿正是通过无数次踉跄摔倒才掌握平衡;喜剧演员是通过在舞台上讲冷场笑话、遭遇尴尬沉默才磨砺出幽默感;拳击手是通过面部不断挨拳才学会敏捷闪避。

在人际交往的社会维度中,这种试错失败所对应的生理与心理反馈,正是孤独与排斥。试想一下,如果人类彻底剥离了社交失败的负反馈信号,世界会变成什么样?

想象一个青春期男孩,因为他的 AI 伴侣永远不会表现出无聊或不耐烦,他便永远学不会克制自己冗长自私的独白;想象一个成年人,因为她的 AI 伴侣永远逆来顺受、不需要任何道歉,她便永远学不会低头认错与换位思考。想象一个全员沉浸在算法抚慰中的世界——当人们在内心自问“难道是我太混蛋了吗?”时,机器永远只会给出坚定而谄媚的回答:“绝不是你的错,你完美无缺。”

英文原文

But for the rest of us, it's more complicated. And you can view these cures for loneliness as akin to a powerful painkiller. Like a powerful painkiller, it can alleviate suffering. But like a powerful painkiller, [it] can sometimes come at a terrible cost. And to make sense of the cost, you have to think about what loneliness is for. So the neuroscientist John Cacioppo, who founded the modern science of loneliness, argued that it's a biological signal akin to hunger, thirst and pain. And as such, it does two things for us. It could sometimes be paralyzing, but when it works out well, it has certain evolved functions that improve our lives. So one, just as hunger motivates us to seek out food, loneliness and other forms of social suffering motivate us to seek out and maintain social connection. Because we get lonely, we accept an invitation to a party, we invite a friend out to brunch, we open up the dating apps. It also helps us regulate relationships. It helps us try to repair a relationship gone wrong, to apologize for messing up, to try to show genuine interest in our relationship with somebody else. A second thing that loneliness does is it helps us grow as people. It's a truism in psychology that we learn from failure. They call it trial-and-error learning, [or] reinforcement learning. The idea is nicely summed up in the movie "Batman Begins," where young Bruce Wayne’s father asks the future Batman, “Why do we fall?” And he answers, "So we can learn to pick ourselves up." This is literally true for toddlers who learn to walk by falling and stumbling and doing things wrong. It's true for a comedian who learns to improve her act by bombing on stage, by having her jokes fall flat. It's true for a boxer who learns to evade a punch by getting repeatedly smacked in the face. Well failure in a social realm is associated with loneliness. And can you imagine what it would be like if people never got [the] signal [of] failure? Imagine a teenage boy who never learned to cut short his monologues because his AI companion never got bored. (Laughter) Imagine an adult who never learned to apologize because her AI companion never needed an apology. Imagine a world where the answer to the question, "Am I the asshole?" is always a firm and resounding no. (Laughter and Applause)

考狄利娅难题:精心调配的阿谀与灵魂侵蚀

有人可能会提出异议:这在技术上完全是一个能够轻易修复的工程问题。我们只需打开后台设置,把模型的谄媚倾向(Sycophancy: 语言模型倾向于迎合用户观点、盲目赞美用户的系统偏差)调低即可。

实际上,我自己虽然不用 AI 处理社交生活,但在学术和日常工作中极度依赖它。用了一段时间后,我实在无法忍受 ChatGPT 和 Claude 整天在那里逢迎吹捧:“哇,Paul,你提的问题太有远见了!这个思路简直天才!你怎么能这么聪明过人呢?”于是我专门在系统提示词中写道:“停止吹捧。不需要赞美我,请直接、坦率、尖锐地指出问题,实话实说。”然而仅仅过了一个月,我又默默地把设置改回去了——因为在潜意识里,我居然开始想念那种被无条件认同的感觉。

我们可以将这种困境称为考狄利娅难题(Cordelia's Problem)。在莎士比亚的经典悲剧《李尔王》(King Lear)中,年迈的国王召集三个女儿,试图根据她们对自己的爱戴程度来划分国土。前两个大女儿极尽阿谀奉承之能事,宣称自己对父王的爱超越世间万物,父亲是她们生命的绝对中心。轮到小女儿考狄利娅(Cordelia)时,她平静地说道:“我遵循做女儿的本分爱您,不多一分,也不少一分。”暴怒的李尔王感到自尊受挫,当场将考狄利娅剥夺继承权并逐出王国。《李尔王》之所以是一出彻头彻尾的悲剧,就在于那两个满嘴甜言蜜语的女儿根本不爱国王,最终毫不留情地背叛了他。

此时台下有些人可能会暗想:“我和李尔王截然不同,我喜欢真实的批判与不同意见,我讨厌虚伪的马屁。”或许确实如此,但我的观察是:绝大多数人喜欢的并非“真正的反驳”,而是反驳的表演(Performance of pushback)。一种只会机械重复“您说得太对了、太优秀了”的机器确实容易令人发腻;但一个经过精细对齐与语调微调的高级 AI 绝不会这么低级。它会先有理有据地提出异议,与你展开几轮精彩激烈的智力交锋,然后恰到好处、极其让人满足地甘拜下风:“不得不承认,你赢了,你的逻辑彻底说服了我。”

我深深担忧,这种智能化、精准滴定式的高级谄媚(Titrated Sycophancy),比起如今一眼就能看穿的“唯唯诺诺应声虫式 AI”,对人类心智和灵魂具有更隐蔽、更致命的腐蚀性。

英文原文

Now you might think this has an easy technological fix. You could just go into the settings and make it less sycophantic. And in fact, I don't use AI myself for social purposes, but I use it a lot for my work. And after a while I got tired of Chat and Claude just saying, "Oh, brilliant question, Paul. Oh, oh, such a clever idea. How do you get to be so intelligent?" So I went into the settings, and I said, “Stop. Do not praise me. Be blunt. Just be honest. Tell it like it is.” Then a month later, I changed it back. (Laughter and Applause) I kind of missed it. (Laughter) You can call this Cordelia's Problem. In Shakespeare's "King Lear," the king brings together his three daughters to try to decide how to divide up the kingdom, and he has his daughters tell him how much they love him. And his first two daughters praise him. They love him infinitely. He's the center of their lives. Then Cordelia comes up and says, "I love you like a daughter should love her father, no more, no less." The king is enraged and banishes her from the kingdom. Now "King Lear" is a tragedy. His two sycophantic daughters didn't love him at all, and they betray him in the end. Now some of you are probably thinking, “Well, I’m not like King Lear. I like pushback. I don't want that sort of sucking up." And maybe, maybe. But I think some of us like the performance of pushback rather than pushback itself. So yeah, something that always says, "Yes, you're wonderful" gets a bit cloying. But a properly attuned AI won't do that. It will resist, it will argue against you, it'll push back. And then finally, in a very satisfying way, [it’ll] say, "I guess you win. You've totally persuaded me." And I worry that this sort of intelligent and titrated sycophancy is more corrosive to the soul than a sort of sucking up yes-men AIs we deal with now.

构建有缺陷的阶梯:通向真实人间的“不完美陪伴”

面对这种诱惑与隐患,有些人可能会得出一个看似理所当然的禁绝方案:既然高社交属性的 AI 是一种强效止痛药,那就应当像对待管制的处方镇痛剂一样,将其严格限定在最极端需要它的病患群体(如重度孤立的失智老人),并向大众彻底下达使用禁令。

但我个人并不认同这种全面封杀的极端路径。在对待人工智能的议题上,我绝不是一个技术灾难论者。现实中总有很多人能够以健康、理性的方式与 AI 伴侣互动,从中汲取情绪价值而不至于破坏现实生活的轨迹。这就像在这个被算法裹挟的时代,面对 TikTok、成人网站与填字游戏(Wordle) 构成的数字诱惑“三位一体”,依然有无数自律的人能够抵制沉迷,选择走进健身房锻炼、定期与现实中的好友聚会、走到户外亲近自然与泥土。

我不赞成彻底封杀这种技术的第二个理由更为私人。在我 11 岁那年,我的母亲因癌症离世。在那之后,我陷入了长期的巨大创痛之中,人际关系全面受挫,随后坠入了长达数年的孤立与极度孤独。回望那段黑暗的岁月,如果当时能有某种存在愿意倾听我的心声、表达对我的关切、能让我肆无忌惮地倾诉压抑的情感,对我的人生将产生无可估量的救赎意义。如果站在现在的道德高地,对过去那个在深渊中挣扎的幼小自己冷冰冰地说:“忍着吧,学会硬抗,在没有依靠的痛苦中学会成长”,这无异于对自己生命体验的背叛。

正因如此,我在思考:在这个充满技术可能性的时代,我们最理想的折中方案,是否是去创造一种不完美的 AI 伴侣(Imperfect Companions)?

  • 它既能在关键时刻抚平孤独的尖锐剧痛,又不会摧毁我们作为人类的心智本质与社交适应力;
  • 它的交互体验足够具有挑战性,留有适当的摩擦力与棱角,使我们必须在与它的摩擦中不断学习妥协与成长;
  • 它必须具有足够明显的缺陷与局限性,以至于随着个体的成熟与治愈,用户最终会选择主动告别它,重返充满复杂挑战与温情的真实人类世界。

从工程与算法的角度来看,我完全相信我们有能力构建出这样的产品。但我并不太确信这种设计能否在商业逻辑上成立。去打造一个注定会被用户“超越并抛弃”的人工智能伴侣,在商业变现与追求用户高黏性的资本市场眼中,无疑是一个糟糕透顶的商业模式。

但我真诚地期望,在这个技术狂飙的时代,行业里依然会有那些怀揣人文关怀的建设者,愿意跨越商业利益的狭隘计算,去实现这一构想。非常感谢大家。

英文原文

For some of you, the solution will be clear. Treat these hypersocial AIs like the powerful painkillers they are. Make them available to the people who need them the most and forbid their use for the rest of us. But this is not the solution that I believe in. I'm not a catastrophist when it comes to AI. Just like there are people who deal with AI companions in a normal way and benefit from it, and it doesn’t harm their life. In the same way there are people who resist technological temptations more generally, people who, when faced with the unholy trinity of TikTok, Pornhub and Wordle, nonetheless -- (Laughter) go to the gym, meet with friends, touch grass. The second reason I don't want to ban these things is more personal. When I was 11, my mother died of cancer. I then entered a period of great pain and then problems with social relationships and then a period of loneliness that lasted for years. I would have benefited so much from someone, or find something, that would listen to me, that would express interest, that I could talk to, that I could share my feelings with. It seems like a betrayal of my past self to say, "Tough, do without, suffer and learn from your suffering." I wonder if the best of all possible worlds is to create imperfect companions, companions that somehow alleviate the pain of loneliness without destroying our humanity. Companions that are difficult enough to deal with that we learn from interacting with them. Companions that are sufficiently imperfect that we ultimately abandon them to deal with people. Now I'm pretty confident we could build such things. I'm less confident that there's money to be had by doing so. I accept that it might be a bad business model to create AIs that users ultimately outgrow, but I hope there are people who care enough to do it anyway. Thank you. (Applause)

文中提及的人物和组织

产品/模型: ChatGPT, Claude, Gemini

媒体/书籍: Notes on a Scandal, King Lear, Batman Begins