AI可能会助纣为虐| 经济学人文化

1

Culture| Johnson

文化 | 约翰逊专栏

英文部分选自经济学人20230722期科技板块

Culture| Johnson

文化 | 约翰逊专栏

AI is making it possible to clone voices

AI正在让“声音克隆”成为现实

That could help fraudsters and disrupt creative industries

这可能会助长诈骗犯的气焰，而且还会干扰创意行业

Jennifer Destefano answered a call from a number she did not recognise. “Mom, I messed up,” her daughter’s voice told her, sobbing. “These bad men have me.” A man proceeded to demand money, or he would drug her daughter and leave her in Mexico. But while she kept him on the phone, friends managed to reach her daughter, only to discover that she was, in fact, free and well on a skiing trip in Arizona. The voice used on the phone was a fake.

詹妮弗·德斯蒂法诺（Jennifer Destefano）曾接到一通陌生来电，电话那头传来了她女儿的抽泣声，“妈妈，我完蛋了，这些坏人绑架了我。”一名男子在电话里开口向她要钱，否则就给她女儿下药，并把她扔在墨西哥。然而，就在德斯蒂法诺拖延着继续和“歹徒”保持通话那会儿，她的朋友们联系上了她的女儿，却发现事实上女儿并未遭遇绑架，而是正在亚利桑那州无忧无虑地享受滑雪之旅。她们这才明白，电话里的声音是经过伪装的。

Ms DeStefano, still shaken, told this story to aUS Senate subcommittee hearing on artificial intelligence in June. The dangers that voice-cloning technology pose are only now starting to be uttered aloud. In recent months, most of the attention paid to artificial intelligence (AI) has gone to so-called “large-language models” like ChatGPT, which churn out text. But voice cloning’s implications will also be profound.

6月，惊魂未定的德斯蒂法诺在美国参议院人工智能小组委员会的听证会上讲述了整件事。时至今日，声音克隆技术所带来的危险才开始为人广泛知悉。近几月以来，各界对AI领域的关注，大多都集中在ChatGPT等可以大量生成文本的所谓“大型语言模型”上。但声音克隆技术的影响也十分深远，不容忽略。

A brief sample of a voice can be used to train an AI model, which can then speak any given text sounding like that person. Apple is expected to include the feature for iPhones in its new operating system, iOS 17, due to be released in September. It is advertised as helping people who may be in danger of losing their voice, for example to a degenerative disease such as ALS.

一份简短的声音样本便可以用来训练AI模型，让它模仿样本说话者的声音说出任何给定的文本。苹果将在9月发布新操作系统iOS17，有可能会在iPhone中添加这一功能。苹果宣传称，该功能可以帮助那些可能面临失声危险的人，比如渐冻症（ALS）等退行性疾病患者。

For those eager to try voice cloning now, ElevenLabs, anAI startup, offers users the chance to create their own clones in minutes. The results are disturbingly accurate. When generating a playback, the system offers a slider that allows users to choose between variability and stability. Select more variability, and the audio will have a lifelike intonation, including pauses and stumbles like “er…” Choose “stability”, and it will come across more like a calm and dispassionate newsreader.

对于那些现在就急切想尝试声音克隆技术的人而言，AI初创公司ElevenLabs就提供了这样一个机会，其用户仅需花上几分钟就能克隆出自己的声音。而且克隆出来的声音逼真得令人不安。在生成回放音频时，系统会为用户提供一个滑动条，两端分别是“多变”和“稳定”。滑动条越接近“多变”一端，音频的语调听起来就越自然，其中还会包含一些停顿和结巴，比如插入“额……”之类的填充词；而滑动条越接近“稳定”一端，声音听起来就越沉稳冷静，就像新闻播音员那样。

Taylor Jones, a linguist and consultant, took a careful look at the quality of ElevenLabs’s clone of his voice in a YouTube video. Using statistical tests he showed that there were a few things off in “his” pronunciation of certain vowels. But a lower-tech test, a “conversation” with his own mother, fooled the woman who raised him. (“Don’t you ever do that again,” she warned.) Johnson repeated the experiment with his own mother, who did not miss a beat in replying to clone-Johnson.

语言学家兼顾问泰勒·琼斯（Taylor Jones）在ElevenLabs平台上克隆了自己的声音，并在一个油管视频中对其音质展开了仔细分析。通过统计检验，他发现克隆声音在某些元音的发音上与他本人有出入。不过，琼斯还做了一个技术含量没那么高的测试，即用克隆声音和他妈妈进行“对话”，并成功骗过了这位从小把他拉扯大的女士。（“你有种敢再试试？”她厉声警告他道。）笔者也做了这一测试，结果发现自己的妈妈与“他的克隆体”交谈顺畅。

For several years, customers have been able to identify themselves over the phone to their bank and other companies using their voice. This was a security upgrade, not a danger. Not even a gifted mimic could fool the detection system. But the advent of cloning will force adaptation, for example by including voice as only one of several identification factors (and thus undercutting the convenience), in order to prevent fraud.

数年来，银行和其他公司可以通过电话里的声音进行客户身份识别。这原本属于安全升级，并不构成危险，因为即便是天才声音模仿家也逃不过检测系统的法耳。然而，由于声音克隆技术的诞生，这些公司将不得不采取应对措施以防范诈骗，比如只将声音作为众多身份识别要素之一（当然，此举会给客户带来不便）。

Creative industries could face disruption too. Voice actors’ skills, trained over a lifetime, can be ripped off in a matter of seconds. TheTelegraph, a British broadsheet, recently reported on actors who had mistakenly signed away rights to their voices, making it possible to clone them for nothing. New contracts will be needed in future. But some actors may, in fact, find cloning congenial. Val Kilmer, who has lost much of his voice to throat cancer, was delighted to have his voice restored for “Top Gun: Maverick”. Others may be spared heading to the studio for retakes. It is the middling professional, not the superstar, who is most threatened.

创意产业也可能受到颠覆性影响。配音演员苦练一生习得的技能，AI只要短短几秒就能掌握。英国《电讯报》近日报道了几位不小心签署合同放弃声音权的演员。这意味着他们的声音可以被免费克隆。未来因此可能需要拟定新合同。然而，在有些演员眼里，声音克隆或许不失为一项友好的技术。方·基默（Val Kilmer）因为喉癌几近失声，而AI在《壮志凌云·独行侠》（Top Gun: Maverick）中再现了他的声音，他为此感到欣喜万分。对于其他演员来说，AI可以免去他们再回演播室重录的麻烦。其实，处境最危险的并不是超级巨星，而是那些不上不下的专业演员。

注释：

1. 创意产业creative industry n.an industry that is based on work in which original ideas are important, such as work in the arts or the media, in designing computer software.

2. rip off v. to perform, achieve, or score quickly or easily

例：ripped off 10 straight points连得十分

3. sign away签字放弃

4. congenial adj. friendly and pleasant友好的，宜人的

5. middling adj. of middle, medium, or moderate size, degree, or quality（大小、程度、品质）中等的，普通的

Another industry that will have to come to grips with the rise of clones is journalism. On-the-sly recordings—such as Donald Trump’s boast of grabbing women by a certain private body part—have long been the stuff of blockbuster scoops. Now who will trust a story based on an audio clip?

除了创意产业，另一个行业也不得不应对声音克隆技术的崛起——新闻业。长期以来，秘密录制内容（比如特朗普炫耀自己抓了女性某私密部位）往往能成为引起轰动的独家新闻。而现在，谁会相信一个基于音频片段的故事呢？

Slightly easier to manage might be the false positives: recordings purporting to be someone but which are fakes. Sophisticated forensic techniques could be of use here, proving a clip to beAI, say, in a courtroom. The opposite problem—the false negatives—will arise when public figures deny authentic recordings. Proving that a clip is genuine is hard, perhaps even impossible. Journalists will need to show how they obtained and stored audio files—unless, as so often, they have promised a source anonymity.

相比之下，假阳性案例或许更容易处理一些，即有些录音声称内容来自某人，实则为假音频。这时，高超的司法鉴定技术或许能派上用场，比如在法庭里证明一段音频为AI所作。当公众人物对真实的录音予以否认时，就会出现相反的问题——假阴性。证明一段音频的真实性并不容易，甚至毫无可能。未来，记者们需要展示他们获取、储存音频文件的方式，除非他们像以往一样，承诺消息来源匿名。

注释：

1. come/get to grips with something: to meet or deal with firmly, frankly, or straightforwardly（坚决地、坦诚地、直接地）面对/解决

2. on the sly: in a manner intended to avoid notice秘密地，私密地

3. scoop n. a story or piece of news discovered and published by one newspaper before all the others独家新闻

4. forensic法医的

5. false negative假阴性 n. a result of a scientific test that appears to show something does not exist or is not present because of imperfect testing methods or procedures, when this is not correct

例：The scans produce false negatives as well as false positives, because they don't catch some lesions.

6. false positive假阳性 n. a result of a scientific test that appears to show something exists or is present because of imperfect testing methods or procedures, when this is not correct

例：The anthrax test is unreliable because it produces too many false positives.

During his first presidential run, Mr Trump did more than anyone to popularise the term “fake news”—and that was well before voice cloning, deepfake videos,artificially generated images and the like were widespread. Now, ever more people caught up in wrongdoing will be tempted by the defence, “It wasn’t me.” Many people will have even more reason to believe them.

特朗普首次参加总统竞选期间，成功带火了“假新闻”一词，其苦心可谓无人能及。当时，声音克隆、深度伪造视频、AI生成图片等还远没有那么普遍。现如今，越来越多的人做错事后都会忍不住自我辩解：“这不是我干的”。而很多人有更充分的理由去相信他们。

注释：

1. run n. 竞选

2. deepfake n. an image or recording that has been convincingly altered and manipulated to misrepresent someone as doing or saying something that was not actually done or said深度伪造的图片/音频/录像

Ryan，译海徜徉的一条咸鱼

Charlie，往者不谏，来者可追

Alfredo，清纯男高体验卡仅剩几个月到期

Alison，人生体验家

YY，愿逆风如解意，容易莫摧残

Constance，痛终有时，爱必将至

3

观点|评论|思考

本周感想

Very，男，电气民工，经济学人资浅爱好者

人们总是倾向于忽视同类所犯的错误，或者习以为常。也许是因为自己也常常会有这种类似的体验：开错车道，或者误听了别人的话，看叉了图像等等。对此人们总是报以相当宽容的态度，但对于AI，或者非同类的错误，人们几乎很难忍受。

人们对于设备的要求很高，有时需要足够的灵敏，能够检测出人们能够感知到的以及不能感知到的一切，有时却要求它具有一定的包容度和柔性，能够以我们的态度为条件来筛选。殊不知这个时候其实已经是过高的要求了，因为我们的态度绝大多数时候是非客观的，或者说我们无法以一套互不矛盾的逻辑体系来表达出来。于是以前的设计人员往往以“做不到”为理由来拒绝这些要求。

但此时，AI的兴起却给了这些常常提出无理要求的人一个可能的机会，因为AI对于他们而言并不是一种科学，而是一种类似同类的对象和工具。在这个假象的基础上，他们便有了充分的理由以自身的感觉和立场来要求实现以往被拒的要求。

在开车时要求车子总能够在遵守交规的前提下在城市拥挤的道路上飞速前行，但是却不能发生任何交通事故。尽管从统计学的基础上来看某些自动驾驶的交通事故率已经低于人类自身，但没人会接受交通事故率不为零的自动驾驶技术，也许当汽车公司承诺包揽一切时，他们才心不甘情不愿地使用这个技术。

我们与生俱来许多缺陷和错误，也许在以前没有一面镜子能够反映出这些问题，但当我们发明的东西可以如此残酷地展现出我们先天的不足时，我们开始害怕了。这是一种基因层面的反应，我们害怕自身会被淘汰，正如我们站在高处会害怕，身处黑洞和暗处会害怕，遇见蛇鼠蜘蛛会害怕，这些是本能，曾让我们在进化的历程中存活了下来。对于AI的害怕同样是这类本能的体现。

也许，是时候到了又一次进化的岔路口。

4

愿景

打造

独立思考 | 国际视野 | 英文学习

小组

01 第十六期写作精品课 

写作课共5位老师
3位剑桥硕士3位博士在读（剑桥，杜伦，港理工）
5位雅思8分（其中两位写作8分，3位写作7.5分）
雅思、学术英语写作，不知如何下笔如有神？
写作精品课带你谋篇布局直播课+批改作文，
带你预习-精读-写作-答疑从输入到输出写出高质量英语作文
点击下图，即可了解写作课详情！

02 早起打卡营 

两年以来，小编已经带着20000多人早起打卡

早起倒逼自己早睡，戒掉夜宵，戒掉手机

让你发现一个全新的自己，创造早睡早起的奇迹！

早起是最简单的自律！

第76期六点早起打卡营

欢迎你的加入！

点击下图，即可了解早起打卡营详情！