Chapter 15. Whose Voice?第十五章 谁的声音?
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 lake, the tapes and the pavilion are in the pictures. Here are the ideas in words.
On pages 1 and 5 Lantern's advert sings in a voice made from somebody who was never asked, and Teo says where such voices come from. Large models learn from enormous collections of text, images, audio and video, much of it gathered from the public internet. Did the people who made all that agree to it? Much of the data was collected without asking. As of 2026 whether that is lawful is contested in courts and parliaments in several countries, and different places are reaching different answers. The ethical point is simpler. Many people did not know, and many feel it should have been their choice. As of 2026 speech systems can also imitate a particular voice from a short recording. That is why the question from Chapter 11, "Did she agree?", matters so much. Meilan could not be asked, and in the story Grandpa answered for her.
On page 10 Teo says the program memorised Grandpa's song. Models mostly learn general patterns, but they can also memorise specific pieces of their training data, and this is more likely for things that are unusual or that appear many times. In 2021 Nicholas Carlini and colleagues extracted hundreds of word-for-word passages from GPT-2, including names, phone numbers and email addresses that had appeared on the web. In 2023 Milad Nasr and colleagues got a production chat assistant to reproduce training text by asking it to repeat a single word forever. These studies are about text. A voice program memorising one family's song is the story's way of showing the same idea. Grandpa's song works as the clue because it is unusual and exists nowhere else.
On page 12 Haoran finds that Willow kept what old people said to it at night. Conversations with chat assistants and recordings from voice assistants may be stored, heard by people who work to improve the system, or used for further training, depending on the company and the settings. In 2023 the US Federal Trade Commission and Department of Justice reached a 25 million dollar settlement with Amazon over allegations that it kept children's Alexa recordings longer than allowed and undermined parents' requests to delete them. Italy's data protection authority also blocked ChatGPT for a month in 2023 over data-protection concerns, and it was allowed back on 28 April after changes.
On page 14 Mira says that teaching Willow again would take a mountain of electricity. AI runs on computers in data centres, and computers use electricity and often water for cooling. Patterson and colleagues estimated in 2021 that training GPT-3 used about 1,287 megawatt-hours and emitted about 552 tonnes of carbon dioxide equivalent. They also stressed that the hardware, the data centre and its location could change the emissions by a factor of 100 to 1,000. The open model BLOOM, trained in France on a largely nuclear-powered grid, was estimated at about 24.7 tonnes for the training energy alone, or about 50.5 tonnes counting the equipment and other overheads (Luccioni, Viguier and Ligozat, 2023). An early, widely quoted paper by Strubell and colleagues (2019) first raised the issue, but its largest figure was later argued to be a big overestimate and should not be quoted as typical. Using a model costs energy too, and it happens billions of times. Generating images takes far more energy than classifying text (Luccioni, Jernite and Strubell, 2024). In 2025 Google reported that its median Gemini text prompt used about 0.24 watt-hours of energy, 0.03 grams of carbon dioxide equivalent and 0.26 millilitres of water. That is less than watching television for nine seconds, but it is a company figure that has not been independently audited, and it leaves out training. For the whole sector, the International Energy Agency estimated that data centres used about 415 terawatt-hours in 2024, about 1.5 percent of the world's electricity, and projected about 945 terawatt-hours by 2030 in its base case. In the United States, the Lawrence Berkeley National Laboratory estimated 176 terawatt-hours, or 4.4 percent, in 2023, and roughly 6.7 to 12 percent by 2028. Li and colleagues estimated that training GPT-3 in Microsoft's US data centres may have evaporated around 700,000 litres of fresh water, a model-based estimate rather than a measurement. So the honest reading is balanced. One question uses very little. The worry is scale, and where the power and water come from. AI may also help, for example with weather forecasts and managing the power grid. The cheap clean power of Bridge City in 2038 is imagined.
On page 18 Teo names a second harm of fakes. In 2019 Robert Chesney and Danielle Citron called it the liar's dividend. When everyone knows that fakes are possible, a person caught on a real recording can say it is fake. Real cases exist now. In January 2024 voters in New Hampshire got robocalls with an AI imitation of President Biden's voice, and the Federal Communications Commission later fined the consultant responsible 6 million dollars. Chapter 11 told the Arup case. The responses are content credentials such as the C2PA standard, invisible watermarks, detection tools that work imperfectly and can be fooled, and law. The European Union's AI Act (Article 50) requires that people be told when they are talking to an AI and that certain AI-generated content be marked, from 2 August 2026. Haoran's promise on page 17, to say whenever a voice is made by a machine, is this idea. The oldest defence still matters most. Check where a recording came from.
One last sentence for families. A voice assistant is not a secret-keeper, so keep names, addresses and private things for people you trust, look at its settings and delete old recordings, ask before you record or copy anyone's voice, and keep a family password for urgent calls, as Xiaoyu and Mama did in Chapter 11.
In the United States the Children's Online Privacy Protection Act requires a parent's consent before online services collect personal information from children under 13. In the European Union, Article 8 of the General Data Protection Regulation sets an age, between 13 and 16 depending on the country, below which a parent must consent for certain online services. UNICEF's guidance on AI and children names data protection and privacy as a core requirement. In September 2025 the US Federal Trade Commission used its study powers to require seven companies that offer AI companion chatbots (Alphabet, Character.AI, Instagram, Meta, OpenAI, Snap and xAI) to report how they test for harms to children and teenagers and how they handle children's data.
Differential privacy adds carefully measured randomness so that results do not reveal any one person's data (Dwork and colleagues, 2006). Data minimisation means collecting only what is needed. Some assistants now run on the device instead of in the cloud. None of these is a complete answer. The most reliable protection for a family is still the simplest, which is not to share private things with a machine.
The 2026 amendment to the EU AI Act adds a ban, from 2 December 2026, on AI systems that generate intimate images of real people without their consent, or sexual images of children.
Bridge City and everyone in it are made up. So is the whole world of 2038, with its garden towers, solar boats and cheap clean power. Lantern, Willow, its dialect launch, its advert, the tool that made the advert and its archive are made up, and so are the made voice in the advert and every word it says or sings. The dialect team, what it gathered, and Willow keeping night recordings and learning from them are made up too. So are Lantern's answers, Mira's agreement, Haoran's live talk, Lantern's promises, Willow's new night voice, and the lab in Zurich and its offer. Grandpa's song "The Moon on the Lake" and Meilan's tapes are the story's, from Chapter 11. The research has measured memorising in text, not in a voice program learning one family's song, and the story uses the song to show the idea. Grandpa's memory is drawn from no one in particular, and the story makes no medical claim. Whether a made voice of someone who has died comforts or hurts is something families weigh for themselves. In this story Meilan could not be asked, and Grandpa answered for her.
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.
湖、磁带和亭子都画在画里了。这里用文字把背后的道理再讲一讲。
第一页和第五页,小灯的广告用一个声音唱歌,可这声音取自一个从没被问过的人。特奥讲了这类声音是从哪儿来的。大模型要靠海量的文字、图片、音频和视频训练,其中很多是从公开网络上收集的。创造这些内容的人同意了吗?很多数据收集时,根本没问过谁。截至2026年,这样做是否合法,好几个国家的法院和议会还在争论,各地得出的结论也不一样。伦理上的道理要简单些。很多人并不知情,也有很多人觉得,这本该由自己决定。截至2026年,语音系统只凭一小段录音,就能模仿某个人的声音。所以第十一章那个问题,“她同意了吗?”,才这样要紧。美兰已经没法回答,故事里是外公替她答的。
第十页,特奥说程序把外公的歌背了下来。模型学到的大多是一般规律,但也会记住训练数据里的具体片段,少见的内容和反复出现的内容尤其容易被记住。2021年,尼古拉斯·卡利尼等人从GPT-2里提取出几百段一字不差的原文,其中有在网上出现过的人名、电话号码和电子邮箱。2023年,米拉德·纳斯尔等人让一个已经上线的聊天助手不停地重复同一个词,结果它吐出了训练文本。这些研究针对的都是文字。故事里让语音程序背下一家人的歌,是借来说明同一个道理。外公的歌能成为线索,正因为它少见,除了这家人的磁带,哪儿都没有。
第十二页,浩然发现小柳存下了老人们夜里对它说的话。跟聊天助手的对话、语音助手的录音,可能会被保存,让改进系统的工作人员听到,或者用于进一步训练,具体要看是哪家公司、怎么设置。2023年,美国联邦贸易委员会和司法部与亚马逊达成2500万美元的和解。两部门指控亚马逊保存儿童Alexa录音的时间超出规定,还阻挠家长的删除请求。同年,意大利数据保护机构出于数据保护方面的担忧,把ChatGPT封了一个月,整改后于4月28日恢复。
第十四页,米拉说,重新教小柳要耗掉一大堆电。AI运行在数据中心的计算机上,计算机要用电,散热往往还要用水。帕特森等人2021年估算,训练GPT-3大约用了1287兆瓦时电,排放了约552吨二氧化碳当量。他们同时强调,硬件、数据中心和选址不同,排放量能相差100到1000倍。开源模型BLOOM在法国训练,那里的电网主要靠核电。据估算,单算训练用电,它排放了约24.7吨;把设备制造等其他开销算进去,约50.5吨(卢乔尼、维吉耶和利戈扎,2023年)。最早提出这个问题的是斯特鲁贝尔等人2019年的论文,流传很广,但后来有人指出,其中最大的那个数字严重高估,不能当作典型情况引用。使用模型同样耗能,而且要用上几十亿次。生成图片的能耗远高于文本分类(卢乔尼、杰尼特和斯特鲁贝尔,2024年)。2025年,谷歌报告称,Gemini处理一条中位数的文本提问,大约耗电0.24瓦时,排放0.03克二氧化碳当量,用水0.26毫升。这比看九秒钟电视还少,但这是公司自报的数字,没有经过独立审计,也不包括训练。放眼整个行业,国际能源署估计,2024年全球数据中心用电约415太瓦时,约占全球用电量的1.5%;按基准情景,到2030年将升至约945太瓦时。在美国,劳伦斯伯克利国家实验室估计,2023年数据中心用电176太瓦时,占全国的4.4%,到2028年约为6.7%到12%。李等人估算,在微软美国数据中心训练GPT-3,可能蒸发了约70万升淡水,不过这是模型推算,不是实测。所以,公允的看法是两头都看。问一个问题,耗电很少。值得担心的是规模,以及电和水从哪里来。AI也可能帮上忙,比如天气预报和电网调度。故事里2038年桥城又便宜又干净的电,是想象出来的。
第十八页,特奥说出了造假的第二重危害。2019年,罗伯特·切斯尼和丹妮尔·西特龙把它称为“说谎者红利”。人人都知道可以造假,被真录音录下的人,就能说录音是假的。现实中已经有这样的事。2024年1月,美国新罕布什尔州的选民接到自动语音电话,里面是AI模仿的拜登总统的声音。联邦通信委员会后来对幕后的顾问罚款600万美元。第十一章讲过阿鲁普公司的案子。应对办法有几种:内容凭证(如C2PA标准)、看不见的水印、并不完美也可能被骗过的检测工具,还有法律。欧盟《人工智能法案》第50条要求,人们和AI对话时要被告知,某些AI生成的内容要加标识,自2026年8月2日起施行。第十七页浩然承诺,凡是机器做的声音都会说明,说的就是这个。最古老的防线仍然最管用:查清录音从哪里来。
最后,给每个家庭一句话。语音助手不会替你保密,名字、住址和私事,只说给信得过的人听;看看它的设置,删掉旧录音;录别人的声音、拷别人的声音之前,先问一声;再像小雨和妈妈在第十一章里那样,定一个家里人遇到急事专用的暗号。
在美国,《儿童在线隐私保护法》规定,在线服务收集13岁以下儿童的个人信息前,必须先征得家长同意。在欧盟,《通用数据保护条例》第8条设了一条年龄线,由各国在13到16岁之间自定。低于这个年龄,使用某些在线服务须经家长同意。联合国儿童基金会关于AI与儿童的指南,把数据保护和隐私列为核心要求。2025年9月,美国联邦贸易委员会动用研究调查权,要求七家提供AI陪伴聊天机器人的公司(Alphabet、Character.AI、Instagram、Meta、OpenAI、Snap和xAI)报告,它们如何检测对儿童和青少年的伤害,又如何处理儿童数据。
差分隐私,是在结果里加入精确控制的随机噪声,让人没法从结果里看出任何一个人的数据(德沃克等人,2006年)。数据最小化,就是只收集确实需要的数据。现在有些助手直接在设备上运行,不上云端。这些办法都不能一劳永逸。对一个家庭来说,最可靠的保护仍是最简单的那一条,私事不要告诉机器。
2026年对欧盟《人工智能法案》的修订新增一项禁令,自2026年12月2日起施行,禁止AI系统未经本人同意生成真人的私密图像,也禁止生成儿童的性图像。
桥城和城里的每一个人都是虚构的,2038年的整个世界也是,包括长满花园的高楼、太阳能小船和便宜的清洁电。小灯、小柳、小柳的方言功能、那条广告、做广告的工具和小灯的资料库,都是编的。广告里那个合成的声音,它说的每句话、唱的每个字,也都是编的。方言组收集了什么,小柳存下夜里的录音再拿来学,同样是编的。小灯的回信、米拉的协议、浩然的直播、小灯的承诺、小柳新的夜间声音,还有苏黎世的实验室和它的邀请,也都是虚构的。外公的歌《湖上的月亮》和美兰的磁带属于这个故事,出自第十一章。研究测过的是模型对文字的记忆,没有测过语音程序记住一家人的歌,故事是借这首歌来说明这个道理。外公的记忆状况不取材于任何具体的人,故事也不作任何医学判断。为已故的人合成声音,是安慰还是伤害,要由每个家庭自己掂量。故事里,美兰没法被问到,是外公替她作了回答。
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。