Chapter 14. Not Everyone Is Heard第十四章 没被听见的人
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 park, the night and the recordings are in the pictures. Here are the ideas in words.
On pages 1 and 5 Willow cannot hear Grandpa, and Teo says why. A machine that learns from examples copies the patterns in them. If the examples are lopsided, so are the patterns, and an unfair world can be repeated at great scale with the look of neutral arithmetic. Researchers describe four ways bias gets in. It can come in through who is in the data, through how the data is labelled, through what the goal measures, and through how the system is used. A speech recogniser learns from recordings paired with what was said. Whisper, for example, learned from 680,000 hours of audio with transcripts gathered from the web (Radford and colleagues, 2022). The research gives one plain example, a voice assistant trained mostly on grown-up speech that struggles with a four-year-old. The story's Bridge City dialect, and Willow's trouble with it, are made up to show this pattern. The research does not report a study of Chinese dialects or of older speakers.
On page 10 the team asks who the other four are. This is the idea of the chapter. In 2018 Joy Buolamwini and Timnit Gebru built a test set of photos of parliament members, balanced by sex and by skin type, and tested three commercial programs that guess a person's sex from a face (from IBM, Microsoft and Face++). All three did best on lighter-skinned men and worst on darker-skinned women. The worst error on darker-skinned women was 34.7 percent, against at most 0.8 percent on lighter-skinned men. The main reason was that the common training and test sets were mostly lighter-skinned faces. A good overall score had hidden it. That is Xiaoyu's question, "who are the other four?", and her move is to count each group on its own.
On page 15 the hospital program counts the wrong thing. In 2019 Ziad Obermeyer and colleagues studied a widely used American program that chose patients for extra care. It used past health spending as a stand-in for need. Less had been spent on Black patients with the same illness, so the program judged them healthier than they were. When the researchers built a version that predicted health more directly, the share of Black patients flagged for extra help rose from 17.7 to 46.5 percent. The program predicted cost correctly. Cost was the wrong thing to predict. The link to the story is that Lantern counted the requests Willow understood, and a person who had given up, like Auntie Zhang, made no requests to count. In 2018 Reuters also reported that an experimental hiring tool at Amazon, trained on ten years of mostly men's CVs, marked down CVs that contained the word "women's" (Dastin, 2018).
On page 17 a public test works. In 2019 Inioluwa Deborah Raji and Joy Buolamwini tested the same companies again after Gender Shades came out, and every audited company had reduced its gap within months. The US National Institute of Standards and Technology ran a much larger study the same year (Grother, Ngan and Hanaoka, 2019). It also found differences between demographic groups for many face recognition programs, and much smaller ones for the most accurate programs.
On page 18 the answer is to report it by group. Margaret Mitchell and colleagues proposed model cards in 2019, short reports of how a model performs for different groups of people and what it should not be used for. Timnit Gebru and colleagues proposed datasheets for datasets in 2021, a record of how a dataset was made, by whom and for what. There are other responses too. People audit and red-team systems, rebalance data or adjust thresholds (each of which builds in a choice), write laws, and keep a person responsible for important decisions. The summary from the experts is that bias cannot be removed once and for all, but it can be measured, reduced and kept visible.
One last sentence for families. If a voice assistant often fails to understand someone you love, the gap is the machine's, not theirs. Tell the maker, and always keep a way to call a person for help that does not depend on being understood by a machine.
In 2016 Julia Angwin and colleagues at ProPublica reported that a score used in American courts to predict who would be arrested again, COMPAS, made different kinds of mistakes for Black and white defendants, and the company that made it replied with a different measure of fairness. The same year Jon Kleinberg, Sendhil Mullainathan and Manish Raghavan, and in 2017 Alexandra Chouldechova, proved that when two groups have different base rates, a score cannot satisfy several natural definitions of fairness at once, except in unrealistic cases. So fairness is a choice about which errors matter most, and that choice is about values, not only maths.
In 2016 Tolga Bolukbasi and colleagues showed that word embeddings had learned "man is to computer programmer as woman is to homemaker", and later work found that the bias stayed hidden in other directions even after the obvious one was removed. In 2017 Aylin Caliskan, Joanna Bryson and Arvind Narayanan found that word embeddings reproduce human implicit associations. In 2023 Federico Bianchi and colleagues found that image generators drew mostly light-skinned men for "a software developer", and that asking for the opposite did not reliably fix it.
The European Union's AI Act (Regulation 2024/1689) treats systems used in hiring, education, credit and policing as high-risk. As of October 2026 a 2026 amending regulation had postponed the main high-risk duties to 2 December 2027 for stand-alone systems and to 2 August 2028 for systems built into regulated products.
Bridge City, its dialect and everyone in it are made up. So is the whole world of 2037, with its garden towers and solar boats. Willow, its voice service, its night music and every word it says are made up, and so are the night phone line and its closing. Lantern, its "ninety-six in a hundred", its pause and its promise to understand Bridge City dialect by spring are made up. The home test, Mr Sun's night (no illness is meant), the hundred speakers, the thousand sentences and every result on the laptop screen are made up too. So are Lili's radio programme, the conference in San Francisco, Ms Ye, Mrs Bai, Mira, Professor Lin's lab and Teo. The research has measured this same pattern with faces, words, images and health and hiring programs, and the studies named in the boxes are real. It does not report a study of Chinese dialects, and the story's numbers about them are invented to show the pattern. The story is set a few years from now and imagines a hopeful future. Grandpa's memory is drawn from no one in particular, and the story makes no medical claim.
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.
公园、那个夜晚和那些录音,都在画里了。背后的道理,这里用文字再讲一遍。
第一页和第五页,小柳听不懂外公,特奥说了原因。机器从例子里学习,就会照搬例子里的规律。例子偏了,学到的规律也跟着偏。一旦大规模使用,现实里的不公平就会被成批复制,表面上却像不偏不倚的计算。研究者总结过偏差进来的四条路:数据里收了谁,数据怎么标注,目标衡量的是什么,系统怎么使用。语音识别模型靠配有文字的录音来学。比如Whisper,用的是从网上收集的68万小时带文字稿的音频(拉德福德等人,2022年)。研究里有个很浅显的例子,一个语音助手主要拿大人的声音训练,听四岁孩子说话就很吃力。故事里的桥城话,以及小柳听不懂桥城话,都是为了说明这条规律编的。我们依据的研究里,没有涉及中国方言或老年人的语音。
第十页,小雨问,剩下那四句是谁说的。这是整章的关键。2018年,乔伊·布奥兰维尼和蒂姆尼特·格布鲁用各国议员的照片组了一套测试集,性别和肤色类型都配平,拿去测三家公司(IBM、微软和旷视)根据人脸判断性别的商用程序。三个程序都是认浅肤色男性最准,认深肤色女性最差。深肤色女性的错误率最高达到34.7%,浅肤色男性最多只有0.8%。主要原因在于,常用的训练集和测试集里大多是浅肤色的面孔。漂亮的总成绩把问题盖住了。小雨问“剩下那四句是谁说的”,问的正是这个。她的办法,就是把每一组单独算一遍。
第十五页,医院用的程序数错了东西。2019年,齐亚德·奥伯迈耶等人研究了美国一个用得很广的程序,它负责挑出需要额外照护的病人。程序拿病人过去的医疗花费,来代表他们需要多少照护。病情一样重的黑人病人,过去花在他们身上的钱少,程序便以为他们比实际更健康。研究者另做了一个更直接预测健康状况的版本,被挑出来接受额外帮助的人里,黑人的比例从17.7%升到了46.5%。原来的程序把花费预测得很准,可花费本来就不该拿来预测。放回故事里,小灯数的是小柳听懂了多少句,而张阿姨那样早就不再开口的人,根本没留下可数的句子。另外,路透社2018年报道过,亚马逊试验过一个招聘工具,拿十年来多半是男性的简历训练,结果给含有“女子”一词的简历扣分(达斯廷,2018年)。
第十七页,公开测试起了作用。《性别阴影》发表以后,伊尼奥卢瓦·黛博拉·拉吉和乔伊·布奥兰维尼在2019年又把那几家公司测了一遍,被点名的公司都在几个月内缩小了差距。同一年,美国国家标准与技术研究院做了一项规模大得多的研究(格罗瑟、恩甘和花冈,2019年)。它同样发现,许多人脸识别程序在不同人群之间有差异,而最准的那些程序,差异要小得多。
第十八页,办法是分组报成绩。玛格丽特·米切尔等人2019年提出了“模型卡”,用一份简短的说明写清模型在不同人群上的表现,以及哪些事不该用它来做。蒂姆尼特·格布鲁等人2021年提出了“数据集说明书”,记下一个数据集由谁做、怎么做、为什么做。应对的办法还有不少:审计系统,请人做“红队”攻击测试,重新配平数据或调整判定门槛(每一种做法都暗含一个取舍),立法,以及让人为要紧的决定负责。专家的总结是,偏差没法一劳永逸地消除,但可以测量,可以减少,也可以一直摆在明处。
最后再叮嘱一句。如果语音助手总听不懂你的家人,毛病出在机器,不在他们。请告诉厂家。另外,一定要给家人留一条能直接找到真人的求助渠道,不靠机器听不听得懂。
2016年,ProPublica的朱莉娅·安格温等人报道,美国法庭用来预测被告会不会再次被捕的评分系统COMPAS,在黑人和白人被告身上犯的错误类型不同。开发它的公司则拿另一种公平标准作了回应。同一年,乔恩·克莱因伯格、森德希尔·穆莱纳坦和马尼什·拉加万证明,两个群体的基础比率不同时,除了一些不现实的特殊情况,一个评分不可能同时满足几种看似都很自然的公平定义。2017年,亚历山德拉·乔尔德乔娃也证明了这一点。所以,公平其实是在选哪种错误更要紧。这是价值观的选择,不只是数学问题。
2016年,托尔加·博卢克巴希等人发现,词向量学会了“男人之于程序员,正如女人之于家庭主妇”这样的类比。后来的研究还发现,就算去掉最明显的那个偏差方向,偏差仍会藏在别的方向里。2017年,艾琳·卡利斯坎、乔安娜·布赖森和阿尔温德·纳拉亚南发现,词向量会重现人类的内隐联想。2023年,费德里科·比安基等人发现,让图像生成程序画“软件开发者”,画出来的多半是浅肤色男性,就算明确要求画相反的人,也不一定管用。
欧盟《人工智能法案》(条例2024/1689)把用于招聘、教育、信贷和警务的系统列为高风险系统。截至2026年10月,2026年的一项修订条例已推迟了高风险系统的主要义务。独立系统推迟到2027年12月2日,嵌入受监管产品的系统推迟到2028年8月2日。
桥城、桥城话和城里所有的人都是虚构的,2037年的整个世界也是,包括长满花园的高楼和太阳能小船。小柳、它的语音服务、它夜里放的音乐和它说的每一句话,都是编的。夜间电话专线和专线要关的事,也是编的。小灯、它的“一百句懂九十六句”、它暂停服务,还有“开春前听懂桥城话”的承诺,都是编的。家里那次测试、孙爷爷的那个夜晚(不指任何一种病)、一百位录音的人、一千句话,还有笔记本电脑屏幕上的每个结果,也都是编的。丽丽的广播节目、旧金山的会议、叶岚、白奶奶、米拉、林教授的实验室和特奥,同样是虚构的。研究者确实在人脸、词语、图像,以及医疗和招聘程序上测出过同样的规律,方框里提到的研究也都是真的。但这些研究没有涉及中国方言,故事里关于方言的数字,是为了说明这条规律编的。故事发生在几年以后,想象的是一个充满希望的未来。外公的记忆问题不取材于哪个具体的人,故事也不作任何医学上的论断。
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。