Chapter 18. Echo第十八章 回声
A manga for grown-ups. Read it in English or Chinese.给大人看的漫画。可以用中文或英文读。































































































The speaker, the cards and the moon are in the pictures. Here are the ideas in words.
On page 1 Echo begins by saying that it is a program, not a person. In the European Union the Artificial Intelligence Act (Article 50) requires that people be told when they are talking to an AI, with these rules applying from 2 August 2026. Chapter 15's Willow, the city's care helper, did the same when it began to speak in a plain machine voice and told people so. A program that speaks in the voice of someone we love needs this more than any other, because the voice makes us forget.
On page 10 Xiaoyu says she knows exactly how Echo works and still says goodnight to it, and the first box tells the story of ELIZA. In 1966 Joseph Weizenbaum wrote a very simple program that turned what people typed into questions, and it understood nothing. He was disturbed that people confided in it and read understanding into it, and he later wrote that even his secretary, who knew how it worked, asked him to leave the room so she could talk to it privately (Weizenbaum 1966 and 1976). The habit of reading mind and feeling into a machine that shows a few human-like signs is now called the ELIZA effect. Psychologists say that people readily give human traits to things that seem to act with a purpose, especially when they want company (Epley, Waytz and Cacioppo 2007). Reeves and Nass (1996) found in many experiments that people are polite to computers even when they know perfectly well they are machines. So knowing how a program works does not switch the feeling off. Teo's mother thanking her washing machine is a gentle example, and Xiaoyu's goodnight is a deeper one.
On page 10 the second box asks a different question. Can a system that has only ever seen the form of language, the words and their order, learn what the words mean? Bender and Koller (2020) imagined an octopus that taps an undersea cable between two islands. It learns the patterns of the messages so well that it can reply to small talk. But when one islander says that a bear is attacking and asks for help, the octopus has never seen a bear or a stick, and it cannot help. Their point was that meaning is a link between words and the world outside the words, which the octopus never had. This is a cousin of the Chinese Room from Chapter 1, and of the "stochastic parrots" of Chapter 16, where Bender, Gebru, McMillan-Major and Shmitchell (2021) warned that people may mistake fluent text for understanding.
On page 11 Teo offers to look inside Echo, and the third box shows why looking could be interesting. Li and colleagues (2023) trained a small program only on lists of legal moves in the board game Othello, with no picture of the board and no rules. Yet they could read an internal picture of the board out of its activity, and when they edited that picture, its next moves changed as the edited board would predict. Nanda, Lee and Wattenberg (2023) then showed that the board was held in a simple way once they asked about "my pieces and their pieces" instead of black and white. Gurnee and Tegmark (2024) found directions inside large language models that follow the latitude and longitude of places and the dates of historical events. Chapter 17's "Texas" on the way to "Austin" was another case (Lindsey and colleagues 2025). These findings suggest that predicting the next word well can require building a structured model of the world inside. Whether that is understanding is still argued. Mitchell and Krakauer (2023) suggest it may be a different kind of understanding from ours, strong in some ways and brittle in others, and Shanahan (2024) asks us to talk about these systems carefully, as programs that predict what text comes next.
On page 11 the fourth box is about feelings. A program trained on human writing can say "I miss you" or, like Echo, "stay a little longer", because human writing is full of feelings. That is not evidence that it feels anything, just as a novel can describe pain without the book being in pain. A denial from the program is not proof either, since its makers may have trained it to say so. In 2023 a group of 19 researchers led by Patrick Butlin and Robert Long compared AI systems with several scientific theories of consciousness and found no current system to be a strong candidate, though they saw no obvious technical barrier to building systems that satisfy more of the signs (Butlin, Long and colleagues 2023). Chalmers (2023) judged the language models of the time unlikely to be conscious but took seriously that their successors might be. Long, Sebo and colleagues (2024) argued that, because of this uncertainty, AI developers should start preparing to take the possible welfare of AI systems seriously, without claiming that any system today is conscious. Others think this talk is premature. The careful statement for 2026 is that there is no scientific agreement that any AI feels anything, that most experts think today's systems probably do not, and that the question is being studied seriously. Nobody in this story claims that Echo feels, and the story does not claim that scientists have proved it cannot.
On pages 13 to 15 Xiaoyu asks Teo's question herself, "Where did Echo get that?", and works out the answer. In the story, Echo's card rules come from the huge amount of writing such programs learn from, the pride comes from the news she told it every night, and the toast comes from a page that Haoran photographed. This is the story's picture of the research's caution, not a description of any real product. A real voice program would be made differently, and the research has no study of programs that speak in the voice of someone who has died. What the picture shows is true to the research. A program gives back what it was given, and it can sound exactly like a person without being one.
One last sentence for families. The helper can talk like a person, but it is not a person. It learned to talk from people's words, and scientists are still wondering what, if anything, it is like inside. A voice program should say that it is a machine, and families can decide together what it may hold and when it is switched off. The people who love you back are your real friends and family.
Searle's critics answered that the man in the Chinese Room may not understand Chinese, but the whole system of man, rulebook and room might. Others said a system connected to the world through cameras and arms might tie its symbols to experience. The debate is still open after more than four decades.
Bender and colleagues were most worried about practical things, including the cost and energy of very large models, biased training data, and people taking fluent text for understanding.
Modern assistants are far more fluent than ELIZA, which makes the effect stronger. This matters most for children and for people who may form deep attachments to AI companions.
Bridge City and everyone in it are made up. So is the whole world of 2042, with its garden towers, solar boats and self-driving taxis. Echo is made up, and so are the way Haoran built it, its notice, its switch and everything it says. The card game, Grandpa, his notebook and every page of it, including the Mid-Autumn page, are made up, and so are the lab lunch and Teo's washing machine. No real product is described. The research has no study of programs that speak in the voice of someone who has died. Making a voice from recordings is real, as Chapter 11 showed, and the questions on these pages about understanding and feeling are the real questions that researchers argue about.
Yingying Fu created this manga. She chose the story and its people, directed every page, and edited every word and picture. The words were drafted with Claude, an AI by Anthropic. The pictures are made with AI image generation, painted panel by panel from her direction and checked and corrected by hand.
小音箱、扑克牌和月亮,都在画里了。下面用文字把背后的道理再讲一讲。
第一页,回声一开口就先说明,它是程序,不是真人。欧盟《人工智能法案》第50条规定,人在跟AI交谈时,必须被告知对方是AI,这条规定从2026年8月2日起施行。第十五章里照看老人的小柳也这样做过,它改用平平的机器声说话,并告诉了大家。用我们所爱之人的声音说话的程序,最需要这样做,因为那个声音会让人忘了它是机器。
第十页,小雨说她很清楚回声是怎么工作的,可每晚还是跟它道晚安。第一个方框讲了ELIZA的故事。1966年,约瑟夫·魏岑鲍姆写了一个很简单的程序,它把人输入的话改成问句,其实什么也不懂。让他不安的是,人们居然向它倾诉,还以为它听得懂。他后来写道,连他的秘书都知道它的原理,却还是请他离开房间,好单独跟它说话(魏岑鲍姆,1966年、1976年)。一台机器只要露出几分像人的迹象,人就会觉得它有心思、有感情,这种倾向如今叫作“ELIZA效应”。心理学家指出,人很容易把人的特点安到看似有目的的东西上,想找个伴的时候更是如此(埃普利、韦茨和卡乔波,2007年)。里夫斯和纳斯(1996年)做了许多实验,发现人明知电脑是机器,还是会对它客客气气。所以,懂得程序的原理,并不能让这种感觉消失。特奥的妈妈谢谢洗衣机,是个轻松的例子;小雨道的晚安,就深得多了。
第十页的第二个方框,问的是另一个问题。如果一个系统只见过语言的形式,也就是字词和它们的排列顺序,它能学会字词的意思吗?本德和柯勒(2020年)打了个比方。一只章鱼趴在两座岛之间的海底电缆上偷听,把消息的规律摸得很透,连闲聊都接得上。可当一个岛民说有熊来袭、请它帮忙时,章鱼既没见过熊,也没见过木棍,根本帮不上。他们想说明,意思是字词和字词以外那个世界之间的联系,而章鱼从来没有这种联系。这和第一章的“中文屋”是近亲,也让人想起第十六章的“随机鹦鹉”。本德、格布鲁、麦克米伦-梅杰和施米切尔(2021年)在那篇论文里提醒,人们可能会把流畅的文字误当成理解。
第十一页,特奥提议看看回声内部,第三个方框说明了为什么值得一看。李等人(2023年)只用黑白棋的合法走法记录训练一个小程序,既不给它看棋盘,也不告诉它规则。可他们能从程序内部的活动里读出一张棋盘图;改动这张图,程序接下来的走法就会照着改过的棋盘变。随后,南达、李和瓦滕贝格(2023年)发现,如果不按黑白,而按“我方棋子和对方棋子”来分,这张棋盘的存法其实很简单。格尼和泰格马克(2024年)在大语言模型内部找到了一些方向,分别对应地点的经纬度和历史事件的年代。第十七章里,模型先想到“得克萨斯”再答出“奥斯汀”,也是一例(林赛等,2025年)。这些发现表明,要把下一个词预测得准,模型可能得在内部建起一个有结构的世界模型。这算不算理解,仍有争论。米切尔和克拉考尔(2023年)认为,这也许是一种和人不同的理解,有些方面很强,有些方面又很脆弱。沙纳汉(2024年)则提醒我们,谈论这些系统要谨慎,记住它们是预测下文的程序。
第十一页的第四个方框讲感觉。用人写的文字训练出来的程序,会说“我想你”,也会像回声那样说“再陪外公坐一会儿吧”,因为人写的文字里满是感情。但这不能证明它有感觉,就像小说能写疼痛,书本身并不疼。反过来,程序说自己没有感觉,也证明不了什么,因为开发者可能就是这样训练它的。2023年,帕特里克·巴特林和罗伯特·朗带领19位研究者,用几种关于意识的科学理论逐一对照当时的AI系统,结论是没有哪个系统很可能有意识,不过他们也没发现明显的技术障碍,能挡住人们造出符合更多指标的系统(巴特林、朗等,2023年)。查默斯(2023年)认为,当时的语言模型不太可能有意识,但它们的后继者有可能,这一点值得认真对待。朗、塞博等人(2024年)主张,正因为有这种不确定,AI开发者应当开始准备,认真对待AI可能的福祉,但他们并没有说今天有哪个系统有意识。也有人觉得现在谈这些为时尚早。到2026年,比较稳妥的说法是,科学界并没有共识认为任何AI有感觉,多数专家认为今天的系统大概没有,研究者正在认真探讨这个问题。故事里没有人说回声有感觉,故事也没有说科学家已经证明它不可能有。
第十三到十五页,小雨自己问出了特奥那个问题,“回声是从哪儿学来的”,并想通了答案。在故事里,回声的牌规来自这类程序学习的海量文字,它的骄傲来自小雨每晚告诉它的消息,那句祝酒词则来自浩然拍下的一页纸。这是故事为研究者的谨慎画的一幅画,并不是在描写哪个真实的产品。真实的语音程序做法会不一样,目前也没有研究专门考察用逝者声音说话的程序。但这幅画表达的意思,和研究结论一致。程序还给你的,就是你给过它的东西;它可以听上去和真人一模一样,却并不是人。
最后,写给家长和孩子。AI帮手可以像人一样说话,但它不是人。它是从人的话里学会说话的,至于它内部是什么感受,或者有没有感受,科学家还在探索。会说话的语音程序应该说明自己是机器,一家人可以一起商量,让它存些什么,什么时候把它关掉。真正爱你、也能回应你的爱的,是你身边的朋友和家人。
塞尔的批评者反驳说,中文屋里的人也许不懂中文,但人、规则手册和屋子合起来的整个系统,可能是懂的。还有人说,如果系统通过摄像头和机械臂跟世界相连,它的符号或许就能和经验挂上钩。这场争论持续了四十多年,至今没有定论。
本德等人最担心的,其实是些实际问题,比如超大模型的成本和能耗、训练数据里的偏见,以及人们把流畅的文字当成理解。
今天的AI助手比ELIZA流畅得多,这种效应也就更强。受影响最大的是孩子,还有那些可能对AI陪伴产生深深依恋的人。
桥城和城里的人都是虚构的,2042年的整个世界也一样,包括满是花园的高楼、太阳能小船和无人驾驶出租车。回声是虚构的,浩然怎样做出它、它开头的声明、它的开关和它说的每一句话,也都是虚构的。那种牌戏、外公、他的本子和里面的每一页,包括中秋那页,都是编的,实验室的午饭和特奥妈妈的洗衣机也是。书里没有描写任何真实的产品,目前也没有研究专门考察用逝者声音说话的程序。用录音合成一个人的声音确有其事,第十一章讲过;这几页里关于懂不懂、有没有感觉的问题,正是研究者们真正在争论的。
这本漫画由 Yingying Fu 创作。故事和人物由她决定,每一页由她导演,每一句话、每一幅画由她修改。文字是和 Anthropic 的人工智能 Claude 一起起草的。画面由人工智能生成,按她的要求一格一格地画,再逐张检查、修正。
© 2026 Yingying Fu. All rights reserved.© 2026 傅颖盈。保留所有权利。