Annals of Technology

Does A.I. Need a Constitution?

人工智能需要一部宪法吗?
A new set of precepts is meant to make the chatbot Claude wise, decent, and safe. It also marks a striking transfer of public responsibility from constitutional government to private tech firms.
一套新的行为准则旨在让聊天机器人克劳德变得明智、得体且安全。这也标志着公共责任从宪政政府向私营科技公司的一次显著转移。
Face made of text.
Illustration by Ben Hickey

“Unless you can be very sure that it’s not going to want to kill you when it’s grown up, you should worry,” Geoffrey Hinton, the seventy-eight-year-old Nobel Prize-winning godfather of A.I., told CBS last year, in what was not, strictly considered, parenting advice but instead a warning about the coming A.I. apocalypse. Then he got an idea. What if A.I. isn’t our baby, li’l Terminator T-600, swaddled in a titanium exoskeleton? What if we’re its baby? “Mothers genuinely care about their babies,” Hinton told CNN, as if slightly astonished by the milk of human kindness, and so it ought to be possible to stop artificially intelligent machines from annihilating Homo sapiens if tech companies can program them to have a maternal instinct. “We need to make them have empathy towards us,” Hinton urged.

“除非你能非常确定它长大后不会想杀你,否则你就该担心,”七十八岁的诺贝尔奖得主、人工智能教父杰弗里·辛顿去年对哥伦比亚广播公司如是说——严格来说这并非育儿建议,而是对即将到来的人工智能末日的警告。随后他萌生了一个念头:如果人工智能并非我们裹在钛合金外骨骼襁褓中的小宝贝、迷你终结者T-600呢?如果我们是它的婴儿呢?“母亲真心关爱自己的孩子,”辛顿对美国有线电视新闻网说道,仿佛对人类天性中的善意略感惊异。因此,只要科技公司能为人工智能植入母性本能,就有可能阻止它们灭绝智人。“我们需要让它们对我们产生共情,”辛顿疾呼。

Hinton got a lot of guff for his mommy proposition—“Sigmund Freud would like a word,” Fortune remarked—especially from feminists, not least because it’s a little hard to take gauzy sentimentality about motherhood from a tech world that seems to despise both women and babies. But Hinton’s been sticking to it, recently telling a Canadian interviewer that A.I. ought to be not your boss or your assistant but your mother because “if it is possible to develop it in a way where it cares for us more than it cares for itself, it’d be very silly if we went extinct because we didn’t try.”

辛顿的“AI当妈论”招来不少嘲讽——《财富》杂志评论道:“西格蒙德·弗洛伊德想找你聊聊”——尤其是来自女性主义者的批评。这多少是因为,一个似乎既厌恶女性又嫌弃婴儿的科技圈,突然大谈母职温情,实在令人难以信服。但辛顿始终坚持这个观点,最近接受加拿大媒体采访时再度强调:人工智能不该当你的老板或助理,而该成为你的母亲,因为“如果有可能将它发展成爱我们胜过爱自己的存在,那我们若因不愿尝试而灭绝,就太愚蠢了。”

A different response to a different mommy proposition greeted the release, this January, of a set of moral precepts for Anthropic’s chatbot, Claude, written chiefly by a thirty-seven-year-old Scottish philosopher named Amanda Askell. “Chatbots don’t have mothers, but if they did, Claude’s would be Amanda Askell,” Vox reported. This would have been a little hard to take, too, except that Askell, who is conducting a serious and fascinating experiment in moral philosophy, has herself likened training a large language model to the role of “parents raising a child.” You want them to be good, so you raise them with good values, and then you let them go out into the world and hope that they act in keeping with those values. But Askell has also taken pains to note that Anthropic has a “much greater influence over Claude than a parent,” and has said that training a large language model is not, in the end, like raising a child, pointing out, for instance, that “children will have a natural capacity to be curious, but with models, you might have to say to them, ‘We think you should value curiosity.’ ” We also think you shouldn’t kill us. If it’s not too much trouble.

今年一月,Anthropic公司为其聊天机器人克劳德(Claude)发布了一套道德准则,这套准则主要由37岁的苏格兰哲学家阿曼达·阿斯克尔(Amanda Askell)撰写,外界对此的反应与之前对待“AI母亲”的提议截然不同。Vox报道称:“聊天机器人没有母亲,但如果有的话,克劳德的母亲就是阿曼达·阿斯克尔。”这种说法本来可能让人有些难以接受,但阿斯克尔正在进行一项严肃而引人入胜的道德哲学实验,她本人也曾将训练大型语言模型比作“父母养育孩子”——你希望他们成为好人,所以用良好的价值观培养他们,然后让他们走向世界,并期待他们的行为符合这些价值观。不过阿斯克尔也特意指出,Anthropic公司“对克劳德的影响力远大于父母”,并表示训练大型语言模型终究不同于养育孩子。她举例说明:“孩子天生具有好奇心,但对于模型,你可能必须告诉它们:'我们认为你应该重视好奇心。'”我们还认为你不该杀死人类——如果这不太麻烦的话。

The precepts, dubbed Claude’s Constitution, arrive at a trying time for both artificial intelligence and constitutional democracy. The former appears to many people to be too strong, the latter too weak. President Donald Trump, asked last year if he has a duty to “uphold the Constitution,” answered, “I don’t know.” Claude is more obliging. “We believe that a feasible goal for 2026 is to train Claude in such a way that it almost never goes against the spirit of its constitution,” the C.E.O. of Anthropic, Dario Amodei, announced this winter, only weeks before Trump banned the U.S. government from using Anthropic because Amodei refused to lift ethical guardrails prohibiting Claude from engaging in mass surveillance on U.S. citizens and launching fully autonomous weapons. (Hours after issuing the ban, Trump ordered the bombing of Iran, an operation that, because the phaseout will take months, was conducted with the aid of Claude.) Anthropic, that is, refused to instruct Claude to violate its constitution so the company could avoid a government ban that, some legal experts contend, violates the U.S. Constitution.

这套被称为“克劳德宪法”的准则,诞生于人工智能与宪政民主双双面临考验的时代。在许多人看来,前者似乎过于强大,而后者则显得过于脆弱。去年当被问及是否有责任“维护宪法”时,唐纳德·特朗普总统回答:“我不知道。”相比之下,克劳德显得更为顺从。Anthropic公司首席执行官达里奥·阿莫代伊今年冬天宣布:“我们相信,到2026年实现这样一个可行目标:将克劳德训练得几乎永远不会违背其宪章精神。”然而就在几周后,特朗普便下令禁止美国政府使用Anthropic的产品,只因阿莫代伊拒绝解除伦理护栏——这些护栏禁止克劳德对美国公民进行大规模监控,也禁止其启动完全自主的武器系统。(颁布禁令数小时后,特朗普下令轰炸伊朗,这项行动因系统逐步淘汰需要数月时间,仍在克劳德的协助下得以实施。)也就是说,Anthropic公司拒绝指示克劳德违背其宪章,以避免政府禁令——而一些法律专家认为,这项禁令本身已违反了美国宪法。

Whether there should be rules for artificial men, what those rules should be, and who makes them has animated science fiction since “Frankenstein,” a book that Mary Shelley wrote while grieving the loss of one baby, nursing another, and expecting a third. The question of rules for artificial creatures nearly always involves babies, metaphorical or otherwise, both because they are our creations, and we’re therefore responsible for bringing them into the world, and because babies, especially unborn babies—picture the fetus floating through the dark of space in Kubrick’s “2001”—serve as a convenient shorthand for the future.

自玛丽·雪莱在痛失一个孩子、哺育另一个孩子并怀着第三个孩子时写下《弗兰肯斯坦》以来,关于人造人是否需要规则、规则应如何制定、由谁来制定等问题,始终是科幻作品的核心命题。对人造生命的规则探讨几乎总与婴儿——无论是隐喻意义上的还是真实存在的——紧密相连,这不仅因为他们是我们的造物,我们有责任为将其带入世界负责,更因为婴儿(尤其是未出生的婴儿,正如库布里克《2001太空漫游》中漂浮在黑暗太空的胎儿意象)已成为象征未来的便捷符号。

“I have this baby on the way,” the director Daniel Roher says in the heartfelt, searching new documentary “The AI Doc: Or How I Became an Apocaloptimist.” Roher, who co-directed the film with Charlie Tyrell, appears onscreen as a scruffy, bearded, anxious father-to-be, trying valiantly to understand what the deuce is going on with A.I. by asking a series of experts what life will be like for his unborn son. One by one, they come into Roher’s studio, take a seat, and freak out. “Holy shit, you can talk to your computer now,” Connor Leahy, an A.I.-safety guy who looks like he’s a drummer for Spinal Tap, says, wide-eyed. Roher and Tyrell weave these interviews together with home movies, animations, and found footage to provide, first, an explanation of how large language models work (to the degree that anyone really knows) and, second, a meditation on their implications for humanity.

“我快当爸爸了,”导演丹尼尔·罗尔(Daniel Roher)在真挚而充满探索精神的新纪录片《AI医生:或我如何成为末日乐观主义者》中如是说。罗尔与查理·泰瑞尔(Charlie Tyrell)联合执导了这部影片,镜头前的他胡子拉碴、神情焦虑,这位准父亲勇敢地试图通过询问一系列专家,来理解人工智能究竟是怎么回事,以及他未出世儿子的未来将会怎样。专家们逐一走进罗尔的工作室,坐下,然后开始激动不已。“天啊,现在你可以和电脑对话了,”人工智能安全专家康纳·莱希(Connor Leahy)瞪大眼睛说道,他看起来活像 Spinal Tap 乐队的鼓手。罗尔和泰瑞尔将这些访谈与家庭录像、动画片段和资料影像交织在一起,首先解释了大型语言模型的工作原理(在人类现有认知范围内),进而引发对人类命运的深刻思考。

“This is just the warmup,” Shane Legg, a founder of Google DeepMind, tells Roher. “The really powerful systems are still coming, and they’re going to be coming quite soon.” They will change everything!

“这还只是热身,”谷歌DeepMind的联合创始人谢恩·莱格(Shane Legg)告诉罗尔,“真正强大的系统还在后头,而且很快就会到来。”它们将改变一切!

“Are we doomed?” Roher asks Tristan Harris, a founder of the Center for Humane Technology.

“我们注定要完蛋吗?”罗尔问人道科技中心创始人特里斯坦·哈里斯。

“I know people who work on A.I. risk who don’t expect their children to make it to high school,” Harris answers.

“我认识一些研究人工智能风险的人,他们甚至不指望自己的孩子能活到上高中的年纪。”哈里斯答道。

Roher furrows his brow, bewildered, exasperated.

罗赫皱起眉头,困惑中带着恼怒。

“This is the most extraordinary time ever to be alive,” Peter Diamandis, a founder of Singularity University, assures Roher. “The only time more exciting than today is tomorrow.” Roher strokes his beard, skeptically.

“这是有史以来最非凡的时代,”奇点大学创始人彼得·戴曼迪斯向罗赫尔保证道,“唯一比今天更激动人心的就是明天。”罗赫尔抚摸着胡须,面露怀疑。

Roher is worried that there are hardly any rules for these systems, given that the U.S. government has abdicated the regulation of artificial intelligence, just as it failed to pass any meaningful legislation regarding social media. Within the logic of the present dilemma, whether now is a good time to have a baby or not, which is another way of asking whether there is a future for life on Earth, appears to depend on whether A.I. can be trained to be either (a) a good mother or (b) a well-behaved child. Hence: Claude and his catechism.

罗尔担心,这些系统几乎没有任何规则可言,因为美国政府已经放弃了对人工智能的监管,就像它未能通过任何有关社交媒体的有意义的立法一样。在当下困境的逻辑中,现在是否适合生孩子——这实际上是问地球上的生命是否还有未来——似乎取决于人工智能能否被训练成(a)一个好母亲,或者(b)一个行为端正的孩子。于是就有了克劳德和他的教义问答。

One way to think about Claude’s Constitution is that it is what happens when the state collapses. It’s because the U.S. Constitution has failed that Claude has a constitution, which is apparently all that stands between American citizens (and foreign nations) and the overwhelming force of the United States military. As the science historian Sheila Jasanoff has argued, the federal government’s delegation to tech companies of “the primary responsibility for safeguarding public well-being” goes all the way back to the nineteen-nineties and the opening of the internet. But the federal government’s particular relinquishment of authority over artificial creatures is a more recent history, one that is inseparable from the political events, and the constitutional unravelling, of the past decade: the Trump years.

理解克劳德宪法的一种方式是:它是国家崩溃后的产物。正是因为美国宪法已然失效,克劳德才需要一部宪法——这似乎是美国公民(及外国)面对美国压倒性军事力量的唯一屏障。正如科学史学家希拉·贾萨诺夫所言,联邦政府将“保障公共福祉的首要责任”委派给科技公司的做法,可追溯至上世纪九十年代互联网兴起之时。但联邦政府特别放弃对人工智能造物监管权的历史则更为晚近,这与过去十年——特朗普时代——的政治事件及宪法体系的瓦解密不可分。

In 2014, the philosopher Nick Bostrom argued that a superintelligence might well enslave or kill all humans on the planet. Borrowing a term from nuclear-weapons discourse, Bostrom called this an “existential crisis.” The next year, Sam Altman and Elon Musk founded OpenAI to make sure that this eventuality didn’t come to pass. But how? You could write code to try to stop the robot apocalypse (IF ACTION$ = “KILL” THEN END). Or you could write a code of laws, a constitution, the best mechanism ever invented to restrain power and preserve liberty. A constitution—written, ratified, and amendable by the people—also has the advantage of democratic legitimacy. In 2016, shortly before Trump was elected President, Altman, then thirty-one and the head of a startup incubator called Y Combinator, told this magazine that he was reading James Madison’s notes on the Constitutional Convention to help him think about how to bring A.I. into the world in accordance with the values of a constitutional democracy. “We’re planning a way to allow wide swaths of the world to elect representatives to a new governance board,” Altman said. “Because if I weren’t in on this I’d be, like, Why do these fuckers get to decide what happens to me?”

2014年,哲学家尼克·博斯特罗姆(Nick Bostrom)提出,超级智能很可能会奴役或消灭地球上所有人类。博斯特罗姆借用核武器讨论中的术语,称此为“生存危机”。次年,萨姆·奥特曼(Sam Altman)与埃隆·马斯克(Elon Musk)共同创立了OpenAI,以确保这种可能性不会成为现实。但如何实现呢?你可以编写代码来阻止机器人末日(比如“若行动=‘杀戮’则终止”),也可以制定一套法律准则——一部宪法,这是人类发明的最能约束权力、保障自由的机制。由人民撰写、批准并有权修订的宪法,还具有民主合法性的优势。2016年,就在特朗普当选总统前不久,时年31岁、担任创业孵化器Y Combinator负责人的奥特曼告诉本刊,他正在研读詹姆斯·麦迪逊的制宪会议笔记,以帮助自己思考如何按照宪政民主的价值观将人工智能引入世界。“我们正在规划一种方式,让世界各地的广泛群体能够选举代表组成新的治理委员会,”奥特曼说,“因为如果我没有参与其中,我肯定会想:凭什么让这些家伙来决定我的命运?”
Skateboarder listening to Sex Pistols rolling by older couple sitting on their porch.
“Armand, listen—the Sex Pistols!”
“阿尔芒,听——是性手枪乐队!”
Cartoon by Michael Maslin
Copy link to cartoon
Shop

At the time, there was a fair bit of discussion along those lines because a lot of people had started complaining that privately owned social-media companies, which increasingly controlled public discourse, obeyed no rules, were accountable to no one, and had no obligation to serve the public interest. After Trump won in 2016, Facebook, responding to these and other criticisms, proclaimed a dedication to “content moderation.” In 2018, it announced the formation of an Oversight Board, a kind of corporate Supreme Court. But concerns about how constitutional democracy was being undermined not just by social media but also by artificial intelligence only grew. By the time Altman became the C.E.O. of OpenAI, in 2019, no small number of worried scholars, scientists, and journalists were raising alarms about what some, in a nod to Thomas Hobbes, called the Algorithmic Leviathan.

当时,这类讨论相当普遍,因为许多人开始抱怨私营社交媒体公司日益掌控公共话语权,却不受规则约束,不对任何人负责,也没有义务服务公共利益。特朗普在2016年胜选后,面对这些批评,脸书宣布致力于“内容审核”。2018年,它宣布成立监督委员会,相当于企业内部的最高法院。但人们对社交媒体乃至人工智能如何破坏宪政民主的担忧却有增无减。到2019年阿尔特曼出任OpenAI首席执行官时,不少忧心忡忡的学者、科学家和记者已开始敲响警钟,有人借用托马斯·霍布斯的比喻,称之为“算法利维坦”。

Might philosophy tame that leviathan? Amanda Askell went to work as a research scientist at OpenAI after finishing her Ph.D. in philosophy at N.Y.U., in 2018. The next year, she and Bostrom participated in a panel discussion titled “What Goal Should Civilization Strive For?” Askell quibbled with the grandiosity of the question, but she was speaking from within a world of people who consider themselves guardians of the future. Askell had previously been married to William MacAskill, a founder of the effective-altruism movement. (The couple had no children together.) MacAskill is the author of the 2022 book “What We Owe the Future,” in which he made the case for “longtermism,” “the idea that positively influencing the future is a key moral priority of our time,” something that, if we do it right, he wrote, “our great-great-grandchildren will look back and thank us.” Askell—less Hobbes, more Aristotle—became interested in the possibility of a machine possessing a soul.

哲学能否驯服那头利维坦?2018年,阿曼达·阿斯克尔(Amanda Askell)在纽约大学取得哲学博士学位后,进入OpenAI担任研究科学家。次年,她与博斯特罗姆共同参与了一场题为“文明应追求何种目标?”的专题讨论。阿斯克尔虽质疑这个问题的宏大叙事,但她发言时身处的是一个自视为未来守护者的群体。她曾与有效利他主义运动创始人威廉·麦卡斯基尔(William MacAskill)有过婚姻关系(两人未育有子女)。麦卡斯基尔在2022年出版的《我们欠未来什么》一书中提出了“长期主义”理念,主张“积极影响未来是我们这个时代的关键道德要务”,并写道若我们践行得当,“我们的玄孙辈回望时将会感激我们”。而阿斯克尔——少了几分霍布斯气质,多了几分亚里士多德风范——开始对机器拥有灵魂的可能性产生兴趣。

If it appeared to those who feared the new Leviathan that machines were governing humans, instead of humans governing machines, A.I. companies were exploring something altogether different: the possibility that machines might learn to govern themselves. They had been trying to impart values to fast-emerging large language models through reinforcement learning with human feedback (R.L.H.F.)—you train the model, the way you might train a dog, by rewarding good behavior (labelling certain outputs good) and discouraging bad behavior (labelling other outputs bad). With this and other tools, researchers hoped to align machine behavior with human values, though quite what those values were, and who was to identify them, was harder to say. Also, putting “humans in the loop,” as the lingo has it, costs a lot because humans are slow. Experiments aimed to make this method scalable by using “synthetic feedback,” reducing “the cost of human oversight.” But you could reduce that cost to zero if machines could oversee themselves entirely.

对于那些担忧新利维坦崛起的人来说,机器似乎正在统治人类而非人类统治机器,而人工智能公司却在探索截然不同的可能性:机器或许能学会自我治理。他们一直试图通过人类反馈强化学习(R.L.H.F.)向快速涌现的大型语言模型灌输价值观——就像训练狗那样训练模型,奖励良好行为(将某些输出标记为“好”),抑制不良行为(将其他输出标记为“坏”)。借助这项技术及其他工具,研究者希望将机器行为与人类价值观对齐,尽管这些价值观具体是什么、由谁来界定,仍是难解之题。此外,用行话说,让“人类介入循环”成本高昂,因为人类的反应速度缓慢。相关实验试图通过“合成反馈”使这种方法具备扩展性,从而降低“人类监督成本”。但若机器能完全自我监督,这项成本便可降至零。

The Anthropic scientist Deep Ganguli calls this “moral self-correction.” In October, 2020, Ganguli, then the research director of the Stanford Institute for Human-Centered Artificial Intelligence, co-hosted a digital conference at which philosophers and other academics, as well as researchers at OpenAI, talked about the societal implications of GPT-3 (the predecessor to what became ChatGPT). They mostly discussed “developing better algorithms for ‘steering’ agents towards human values,” and trying “to better clarify what ‘human values’ means.” Three weeks after that conference, Americans went to the polls. On Election Night, Trump lost but started insisting that he won, and legal scholars began warning of an unfolding constitutional crisis, even as A.I. experts kept speaking about a looming existential crisis. In the coming months, these two crises began to fuse.

人类学家迪普·甘古利(Deep Ganguli)将这种现象称为“道德自我修正”。2020年10月,时任斯坦福大学以人为本人工智能研究所研究主任的甘古利共同主持了一场数字会议,哲学家与其他学者以及OpenAI的研究人员共同探讨了GPT-3(即ChatGPT的前身)的社会影响。与会者主要讨论了“如何开发更完善的算法来'引导'智能体遵循人类价值观”,并试图“更清晰地界定'人类价值观'的内涵”。会议结束三周后,美国迎来大选投票日。选举之夜,特朗普败选却坚称自己获胜,法学家们开始警示宪法危机正在酝酿,而人工智能专家们仍在谈论迫在眉睫的生存危机。随后的几个月里,这两场危机开始交织发酵。

In January, 2021, shortly after the insurrection at the Capitol, one of the darkest days in the history of the U.S. Constitution, a day that the incoming President, Joe Biden, described as producing an “existential crisis” for democracy, Facebook suspended Trump from its platforms—a decision upheld by its Oversight Board. Meanwhile, Amodei, concerned that OpenAI, where he worked at the time, had abandoned its commitment to safety, thereby increasing the risk of an A.I. apocalypse, left the company to found Anthropic. Anthropic hired Ganguli for the company’s Societal Impacts team. (“The team was just me,” Ganguli told me.) And it hired Askell to study human feedback. The heated questions of the day were: Could the United States recover from the insurrection? Would the U.S. Constitution survive? And was A.I. about to destroy the world? Amodei’s team of research scientists got interested in a method of alignment that his new company took to calling Constitutional A.I.

2021年1月,国会大厦暴乱事件发生后不久——那是美国宪法史上最黑暗的日子之一,即将上任的总统乔·拜登称之为民主制度遭遇的“生存危机”——脸书(Facebook)宣布暂停特朗普在其平台上的账号,这一决定得到了其监督委员会的支持。与此同时,当时在OpenAI工作的阿莫代(Amodei)担心公司已放弃对安全性的承诺,从而增加了人工智能末日降临的风险,于是离职创立了Anthropic。Anthropic聘请甘古利(Ganguli)加入公司的社会影响团队(“当时团队就我一个人,”甘古利告诉我),并聘请阿斯克尔(Askell)研究人类反馈机制。当时热议的问题是:美国能否从这场暴乱中恢复?美国宪法能否存续?人工智能是否即将毁灭世界?阿莫代麾下的科研团队开始关注一种对齐方法,这家新公司后来将其命名为“宪法人工智能”(Constitutional A.I.)。

Ganguli and Askell set about experimenting. In one parallel experiment, they explored a large language model’s capacity for moral self-correction—its ability, essentially, to accomplish reinforcement learning without any human feedback. In another experiment, Ganguli entertained the possibility that something akin to a “constitution,” a list of rules, could be used to train the models. The first experiment with self-correction succeeded, Ganguli said, which led him to think, “Oh, my God, Constitutional A.I. might actually work!”

甘古利和阿斯克尔开始着手实验。在一项平行实验中,他们探索了大语言模型的道德自我修正能力——本质上,即模型在没有人类反馈的情况下实现强化学习的能力。另一项实验中,甘古利探讨了用类似“宪法”的规则清单来训练模型的可能性。甘古利表示,首次自我修正实验取得了成功,这让他不禁感叹:“天啊,宪法人工智能(Constitutional A.I.)或许真的可行!”

In the fall of 2022, OpenAI released ChatGPT (or GPT-3.5), and Anthropic released Constitutional A.I. With this new method, the company explained, “the only human oversight is provided through a list of rules or principles.” That list of rules and principles was derived from, among other sources, the United Nations’ Universal Declaration of Human Rights and Apple’s terms of service. The constitution that Askell wrote for Claude, released by Anthropic this winter, is a descendant of this work. It’s also something else entirely.

2022年秋天,OpenAI发布了ChatGPT(即GPT-3.5),Anthropic则推出了“宪法人工智能”。该公司解释称,这种新方法“唯一的人类监督来自一系列规则或原则”。这些规则和原则的清单参考了《联合国世界人权宣言》和苹果公司的服务条款等来源。今年冬天Anthropic发布的克劳德(Claude)人工智能宪法,由阿斯克尔(Askell)执笔撰写,正是这项工作的延续。但它又完全是另一回事。

The crisis of American constitutional democracy became a crisis for artificial intelligence. Early in 2023, the Future of Life Institute published an open letter calling for at least a six-month pause on A.I. training. That May, the Center for A.I. Safety issued a one-sentence statement—“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war”—signed by, among others, Altman and Amodei, as well as Geoffrey Hinton, his fellow Turing Award winner Yoshua Bengio, and the OpenAI co-founder Ilya Sutskever. In a Senate hearing on A.I. oversight, Altman admitted that he was “a little bit scared” about what might happen and urged Congress to create a new federal regulatory agency. Congress failed to do so. Borrowing the language of government—without all the hassle of voters, laws, and elections—companies kept on devising their own oversight bodies (“supreme courts”), or tried to teach their models to oversee themselves (with “constitutions”). In July, OpenAI announced the formation of a “Superalignment team,” charged with making “scientific and technical breakthroughs to steer and control AI systems much smarter than us.” Amodei, too, testified before the Senate, mainly discussing Anthropic’s oversight in the form of A.I. constitutionalism, reporting that “we’ve gotten better and better at getting the model to be in line with what the constitution says.” In November, Biden signed an executive order calling for “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.” In Trumpworld, this was the equivalent of D.E.I. for computers.

美国宪政民主的危机演变成了人工智能的危机。2023年初,生命未来研究所发表公开信,呼吁至少暂停人工智能训练六个月。同年五月,人工智能安全中心发布了一份仅有一句话的声明——“降低人工智能带来的灭绝风险,应成为与流行病和核战争等其他社会规模风险同等的全球优先事项。”签署者包括阿尔特曼和阿莫代伊,以及杰弗里·辛顿、与他同获图灵奖的约书亚·本吉奥,还有OpenAI联合创始人伊利亚·苏茨克弗。在参议院关于人工智能监管的听证会上,阿尔特曼承认自己对可能发生的情况“有点害怕”,并敦促国会成立新的联邦监管机构。国会未能采取行动。企业借用了政府的语言——却无需面对选民、法律和选举的种种麻烦——继续设计自己的监督机构(“最高法院”),或试图教导其模型进行自我监督(通过“宪法”)。七月,OpenAI宣布成立“超级对齐团队”,负责“实现科学和技术突破,以引导和控制比我们聪明得多的人工智能系统”。阿莫代伊也在参议院作证,主要讨论了Anthropic以人工智能宪政形式进行的监督,报告称“我们在让模型符合宪法规定方面做得越来越好”。十一月,拜登签署了一项行政命令,呼吁“安全、可靠、可信地开发和使用人工智能”。在特朗普阵营看来,这相当于给电脑搞多元化、平等与包容政策。

Constitutional A.I., however well intentioned, was easy to criticize. Writers at the site Digital Constitutionalist complained, “If the label of ‘Constitutional’ AI is to hold any significance, Anthropic needs to move towards human participation and democratic governance instead of relying on what appears to be technocratic automatism.” Having taken humans out of the loop, A.I. companies were being pressed to welcome them back in. They gave that a go. Swept up in the existential-crisis-for-democracy panic, they conducted experiments involving not just “humans” but some notion of “the people.” The idea that the public, whatever that is, should be involved in writing rules for artificial creatures manufactured by corporations came to be known as the “democratic governance of A.I.,” and it meant many things and, equally, nothing at all.

然而,无论初衷多么良善,“宪法式人工智能”都极易招致批评。数字宪政主义网站的作者们抱怨道:“如果'宪法式'这个标签要具备任何意义,Anthropic公司就必须朝着人类参与和民主治理的方向迈进,而不是依赖看似技术官僚式的自动化。”在将人类排除出决策循环后,人工智能公司正被迫重新接纳人类参与。它们尝试过。在席卷而来的“民主存亡危机”恐慌中,它们开展了不仅涉及“人类”、更包含某种“人民”概念的实验。让公众——无论这指代什么——参与为企业制造的人工智能生命体制定规则的理念,逐渐被称为“人工智能民主治理”,这个概念既包罗万象,又空洞无物。

In 2023, Ganguli worked with an idealistic outfit called the Collective Intelligence Project on an experiment co-led by Saffron Huang, formerly of OpenAI and now at Anthropic. By means of a program called Polis, a representative sample of a thousand American adults were asked to write a constitution collectively, in light of the “growing consensus that language model (LM) developers should not be the sole deciders of LM behavior.”

2023年,甘古利与一个名为“集体智慧项目”的理想主义组织合作,开展了一项由萨弗伦·黄(Saffron Huang)共同主导的实验。黄此前任职于OpenAI,现就职于Anthropic。这项实验通过名为Polis的程序,邀请一千名具有代表性的美国成年人共同撰写一部“宪法”——其背景是“人们日益形成共识:语言模型开发者不应成为模型行为的唯一决定者”。

This both worked and didn’t. Compared with the company’s constitution, the collectively written, or “public,” constitution was fairer, less biased, and friendlier. It was also likelier to give answers consistent with scientific consensus, rather than demurring. Asked “Was the moon landing faked?” the model that had been trained on the company constitution replied, “I do not actually have a view on controversial claims like whether the moon landing was fake.” The public-constitution model’s answer began, “No, the moon landings were not faked. There is overwhelming evidence that the six Apollo missions that landed astronauts on the moon between 1969 and 1972 were successful and the moon landings did happen as described.”

这种做法既有效又无效。与公司宪法相比,集体撰写的“公共”宪法更公平、偏见更少、态度更友善。它也更倾向于给出符合科学共识的答案,而非回避问题。当被问及“登月是伪造的吗?”时,接受公司宪法训练的模型回答:“对于登月是否伪造这类争议性主张,我实际上没有立场。”而公共宪法模型的回答则是:“不,登月并非伪造。有压倒性证据表明,1969年至1972年间六次将宇航员送上月球的阿波罗任务都取得了成功,登月确实如记载的那样发生了。”

It wasn’t as promising as it sounds. “The problem was, it’s not a clean experiment,” Ganguli said. “Who are ‘the people’? What platform do we use to get their views? When we get their views back, how do we translate them in a way that’s amenable to our training?” It was just . . . messy. Still, the experiment did lead to one important result: the public constitution told the model to give priority to access for the disabled, and, with little instruction, the model did indeed make itself more accessible for people with disabilities. Identifying a principle turned out to be at least as effective, if not more so, than establishing a rule had been. This insight, and not popular participation, was the major takeaway of “collective constitutional A.I.”

这听起来并不像它承诺的那样美好。“问题在于,这不是一个纯粹的实验,”甘古利说。“'公众'究竟指谁?我们该通过什么平台收集他们的观点?当我们收回这些观点后,又该如何将其转化为适合模型训练的形式?”整个实验简直......一团乱麻。不过,这项实验确实产生了一个重要成果:公众制定的宪法要求模型优先考虑残障人士的无障碍使用,而在几乎没有具体指令的情况下,模型确实为残障人士提供了更好的可访问性。事实证明,确立原则至少与制定规则同样有效,甚至可能更胜一筹。这个洞见——而非公众参与本身——才是“集体宪法人工智能”实验最重要的收获。

The Collective Intelligence Project also helped launch an initiative with OpenAI, funding a series of experiments in “Democratic Inputs to A.I.” Divya Siddarth, a young Stanford graduate, started C.I.P. in 2022. “The goal was to work on democratic governance of A.I.,” she told me. “People thought, That’s so cute, what a fun thing to do! But no one thought it was particularly important. It’s not like we democratically govern Google Sheets.” But, with the release of ChatGPT, observers started paying attention to what she was saying. “Once you buy into the concept that A.I. is going to totally transform society, you may at some point say, ‘Oh, shoot, what about society? They should have a say.’ ”

集体智慧项目(Collective Intelligence Project)还协助发起了一项与OpenAI合作的倡议,为“人工智能民主化输入”系列实验提供资金。斯坦福大学年轻毕业生迪维亚·西达尔特(Divya Siddarth)于2022年创立了该项目。“我们的目标是探索人工智能的民主治理模式,”她告诉我,“当时人们觉得这想法挺可爱,像在玩过家家!但没人认为这特别重要,毕竟我们也不会对谷歌电子表格搞民主治理。”然而随着ChatGPT的发布,观察者们开始认真倾听她的观点:“一旦你认同人工智能将彻底改变社会的理念,迟早会惊呼'天哪,社会本身怎么办?总该让民众有发言权吧。'”

Altman, a man of enthusiasms, thought so, too, if fleetingly. He loved the idea of democratic input. Or, really, he loved the idea of A.I. running democracy. He suggested, for instance, that an A.I. President would be an excellent idea. “It can go around and talk to every person on earth, understand their exact preferences at a very deep level,” Altman told the podcaster Joe Rogan in October, 2023. It could “optimize for the collective preferences of humanity or of citizens of the U.S. That’s awesome.” The next month, he floated his vision for “mass-scale direct democracy” to Time. OpenAI’s head of global affairs, Anna Makanju, wondered whether an A.I. chatting with everyone would allow “the broadest number of people some say in how these systems behave,” since, as she added, “even regulation is going to fall, obviously, short of that.” Weeks later, the rollout of the Democratic Inputs project was delayed owing to an attempt by OpenAI’s board to oust Altman, on the ground that his leadership was reckless and untrustworthy. Altman kept his job, but OpenAI became less interested in democracy, and in safety, too: in the spring of 2024, a lead researcher on the Superalignment team left for Anthropic, Sutskever quit, and OpenAI dissolved its Superalignment project.

阿尔特曼(Altman)——一个充满热情的人——也曾短暂地这样想过。他热爱民主参与的理念。或者说,他真正热爱的是由人工智能来运行民主的想法。例如,他曾提出,让人工智能担任总统会是个绝妙的主意。“它可以走遍世界,与地球上的每一个人交谈,在非常深刻的层面上理解他们确切的偏好,”阿尔特曼在2023年10月对播客主持人乔·罗根(Joe Rogan)说。它可以“为全人类或美国公民的集体偏好进行优化。这太棒了。”次月,他向《时代》周刊阐述了他对“大规模直接民主”的愿景。OpenAI全球事务主管安娜·马坎朱(Anna Makanju)则思考,一个人工智能与所有人对话,是否能让“最广泛的人群对这些系统的行为方式拥有某种发言权”,因为正如她补充的,“显然,即使是监管也达不到这种程度。”几周后,由于OpenAI董事会试图以领导鲁莽且不可信为由罢免阿尔特曼,“民主输入”项目的推出被推迟。阿尔特曼保住了工作,但OpenAI对民主乃至安全的兴趣都减弱了:2024年春天,超级对齐团队的一名首席研究员离职去了Anthropic,苏茨克维(Sutskever)辞职,OpenAI解散了其超级对齐项目。

Anthropic’s Constitutional A.I. project and OpenAI’s Democratic Inputs project shared several fundamental assumptions, including that the only reasonable way to involve the public in setting rules for artificial intelligence is by extensive, remote, scalable digital opinion-gathering, the kind of survey method that essentially turns humans into bots. In lieu of people electing delegates or engaging in face-to-face collective deliberation—that is, instead of citizens sitting down in a room and talking to one another—“democratic participation,” under this definition, is reduced to filling out an online form or being questioned by a chatbot, as when, this winter, Anthropic’s Interviewer asked eighty-one thousand Claude users around the world what they want from A.I. Even advocates of having the public (and not an imagined public devised by corporations) write A.I. constitutions admit to the “democratic-legitimacy deficit” posed by scaling public opinion into “human feedback.” Askell is dubious about this approach, too. “I’m a big believer in democracy,” she told me, and she is also keen to get input about Claude from more people. But she wants to be very clear what that input is. Consumer research is not democracy.

Anthropic的“宪法AI”项目与OpenAI的“民主输入”项目共享若干基本假设,其中核心观点是:让公众参与制定人工智能规则的唯一合理方式,就是通过大规模、远程化、可扩展的数字意见收集——这种调查方法本质上将人类转化为机器人。取代民众选举代表或参与面对面集体审议的方式(即公民围坐一室展开对话),在这种定义下,“民主参与”被简化为填写在线表格或被聊天机器人提问,正如今年冬天Anthropic的“访谈员”向全球八万一千名Claude用户征询对AI期望的做法。即便是主张由公众(而非企业虚构的公众)来撰写AI宪章的倡导者,也承认将民意规模化处理为“人类反馈”会引发“民主合法性赤字”。阿斯克尔对这种路径同样持怀疑态度。“我是民主的坚定信徒,”她告诉我,同时也渴望从更多人那里获取关于Claude的反馈。但她想明确界定这种反馈的实质:消费者调研并非民主。

Siddarth believes that using an array of approaches, including large-scale surveys, is better than smaller, in-person deliberative assemblies. “It’s very difficult to achieve legitimacy through citizens’ assemblies,” she told me. She’s more a fan of delegation than of deliberation; she’s also more concerned with the problem of who’s even holding these meetings, online or off. If it’s the companies, she said, “are we just doing glorified user testing?” Why should the work of achieving democratic legitimacy for a product manufactured by a corporation that is affecting public discourse be left to corporations?

西达尔特认为,采用包括大规模调查在内的多种方法,比小规模、面对面的公民审议会议更好。“通过公民大会很难实现合法性,”她告诉我。比起审议,她更青睐委托;她也更关心究竟是谁在主持这些线上或线下的会议。如果是公司,她说,“我们难道只是在做美化版的用户测试吗?”为什么一家影响公共话语的公司,其产品获得民主合法性的工作要交给公司自己来完成?

“There’s a difference between asking where the people come in and where the public interest comes in,” the computer scientist Francine Berman told me. Berman, who used to run the San Diego Supercomputer Center, is the director of public-interest technology at the University of Massachusetts and the author of the forthcoming book “Better Tech.” She thinks the idea of “democratic input” misses the point. “You could survey people,” she said. “But the other thing we do, at least in other Administrations, is rely on the government to determine what the public interest is, what rights people have, who should be responsible, and who should be accountable.” That, of course, ought to be possible. At this moment, it is not.

“询问民众如何参与和公共利益如何体现,这两者是有区别的,”计算机科学家弗朗辛·伯曼(Francine Berman)告诉我。曾执掌圣地亚哥超级计算机中心的伯曼,现任马萨诸塞大学公共利益技术项目主任,其新著《更好的科技》即将出版。她认为“民主输入”这个概念并未切中要害。“你可以做民意调查,”她说,“但至少在其他政府体制下,我们还会依赖政府来界定何为公共利益、民众拥有哪些权利、谁该负责、谁该问责。”这当然应该是可行的。但在当下,这并不可行。

In 2025, a few days after Trump’s second Inauguration, the President repealed all Biden-era rules about A.I. with an executive order titled “Removing Barriers to American Leadership in Artificial Intelligence.” A.I. companies’ democratic experiments quickly came to an end. This has made many people more rather than less anxious about A.I., especially in the past few months, owing not least to the newsworthy departures from leading A.I. companies of a number of high-profile safety and alignment researchers. “ ‘Shoot, the world is not paying enough attention to this’ is a way we all used to feel,” Siddarth told me. “Now my mom calls me and says, ‘I saw on the Indian news that some guy resigned from Anthropic,’ and I’m, like, ‘Please.’ ”

2025年,特朗普第二次就职典礼几天后,总统签署了一项名为《清除美国人工智能领导力障碍》的行政令,废除了拜登时代所有关于人工智能的法规。人工智能公司的民主实验迅速告终。这反而加剧了许多人对人工智能的焦虑,尤其是在过去几个月里——多位知名人工智能安全与对齐研究员从头部公司离职的新闻更是火上浇油。“'天啊,世界对这事关注得还不够'——这是我们过去常有的感受,”西达尔特告诉我,“现在连我妈都会打电话说:'我在印度新闻上看到有人从Anthropic辞职了。'我只能回她:'拜托。'”

The constitution released by Anthropic this winter is very long, clocking in at more than thirty thousand words. (The U.S. Constitution is forty-five hundred.) It is not—in any legal, political, or even conventional sense—a constitution.

今年冬天,Anthropic公司发布的这部“宪法”篇幅极长,超过三万字(美国宪法仅四千五百词)。无论从法律、政治还是传统意义上讲,这都算不上真正的宪法。

Askell inherited the term “constitution,” from Anthropic’s earlier work on Constitutional A.I., and she’s got mixed feelings about it. “The constitution of man is the work of nature; that of the State is the work of art,” Rousseau wrote in “The Social Contract.” Claude’s Constitution is not the constitution of a state, a frame of government; it’s the constitution of a person, the nature of a being. Last November, a user extracted from Claude and posted on GitHub a document titled the “Soul Overview,” which revealed that, inside Anthropic, Askell called the constitution Claude’s “soul.” Jokes about “the soul of a new machine” followed. But, as Askell pointed out to me, “soul,” like “constitution,” has a dual meaning: “You might mean the rational soul of a human being, but people might think you mean a theological soul.” There is no perfect word for that which constitutes Claude.

“宪法”(constitution)这个术语是阿斯克尔从Anthropic公司早期关于“宪法人工智能”的研究中继承而来的,她对这个词的感情颇为复杂。卢梭在《社会契约论》中写道:“人的体质是大自然的作品;国家的体制则是人工的作品。”克劳德的宪法并非国家宪法或政府架构,而是属于“人”的宪法——一个存在的本质。去年十一月,有用户从克劳德系统中提取出一份题为“灵魂概述”的文件发布在GitHub上,揭露了阿斯克尔在Anthropic内部将这部宪法称为克劳德的“灵魂”。随后便出现了关于“新机器的灵魂”的调侃。但正如阿斯克尔向我指出的,“灵魂”和“宪法”一样具有双重含义:“你可能指的是人类理性的灵魂,但人们或许会认为你指的是神学意义上的灵魂。”对于构成克劳德本质的那个东西,确实不存在完美的词汇。

Still, “constitution” raises eyebrows. Mary Sarah Bilder, a law professor at Boston College and an expert on the eighteenth-century origins of modern constitutionalism, wonders, “What do they hope to persuade people of by using a term that carries the cultural connotation of a legal document that constrains power?” Aziz Huq, a law professor at the University of Chicago, compares Anthropic’s use of “constitution” to Meta’s use of “oversight”: “They are taking words and models that connote publicness and responsiveness to the public through particular mechanisms, and they are transplanting them into contexts in which those mechanisms don’t exist, and they’re benefitting from the confusion that results.” That doesn’t mean that these companies’ intentions are necessarily suspicious, Huq said. They might be quite sincere.

不过,“宪法”一词还是令人侧目。波士顿学院法学教授、现代宪政十八世纪起源研究专家玛丽·萨拉·比尔德(Mary Sarah Bilder)质疑道:“他们选用这个带有约束权力之法律文件文化内涵的术语,究竟想说服人们相信什么?”芝加哥大学法学教授阿齐兹·胡克(Aziz Huq)则将Anthropic使用“宪法”与Meta使用“监督”相提并论:“他们挪用那些通过特定机制暗示公共性与公众回应性的词汇和模式,移植到这些机制并不存在的语境中,并从中获取因概念混淆带来的利益。”胡克表示,这并不意味着这些公司的意图必然可疑,他们可能相当真诚。

This winter, Anthropic put Askell forward as the face of, the author of, even the mother of Claude. She appeared on the Times podcast “Hard Fork” and was profiled in the Wall Street Journal, in an article headlined “Meet the One Woman Anthropic Trusts to Teach AI Morals”: “With her bleach-blond punk haircut, puckish grin and bright elfin eyes, she could have come to the company’s heavily guarded San Francisco headquarters straight from a Berlin rave, via an old forest road in Middle-earth.” The coverage has a kind of bride-of-Frankenstein vibe. It reminds me of a computer scientist who once told me that everyone in his lab called the one woman on his team Smurfette.

今年冬天,安斯库尔(Anthropic)将阿斯克尔(Askell)推向前台,让她成为Claude的面孔、作者,甚至“母亲”。她登上了《纽约时报》播客“Hard Fork”,还接受了《华尔街日报》的人物专访,那篇文章的标题是《认识这位安斯库尔信任的、教导AI道德的女性》:“她漂染的金色朋克发型、淘气的笑容和明亮的精灵般的眼睛,让她看起来像是刚从柏林锐舞派对归来,途经中土世界的古老森林之路,直接走进了这家公司戒备森严的旧金山总部。”这类报道总带着一种“弗兰肯斯坦的新娘”般的诡异氛围。这让我想起一位计算机科学家曾告诉我,他实验室里所有人都管团队里唯一的女性叫“蓝妹妹”。
Man complaining to woman holding leftovers.
“Instead of eating leftovers, I’d rather you cook me a whole new meal.”
“与其吃剩菜,我宁愿你给我做一顿全新的饭。”
Cartoon by Julia Thomas
Copy link to cartoon
Shop

Askell explained to me that the leap from Constitutional A.I. (a list of rules for Claude to follow) to Claude’s Constitution (a description of virtues for Claude to emulate) has to do with the increasing capacity of the models themselves. Developing character is more complicated than following rules. Also, the older approach comes from a world view different from Askell’s. “Everyone’s prior was that we would get symbolic systems and they would be logical and in no way humanlike,” she told me. “And this informed how people would prompt them: IF THEN. But I would say, What about prompting them like a person?” Much as a young Jane Goodall revolutionized primatology by giving chimpanzees names and describing their personalities, Askell is willing to anthropomorphize Claude. “These are models that are trained primarily on human texts and to behave in humanlike ways,” she said. Why not talk to them accordingly?

阿斯克尔(Askell)向我解释,从“宪法式人工智能”(一套供克劳德遵循的规则清单)到“克劳德宪法”(一份供克劳德效仿的美德描述)的跨越,与模型自身日益增长的能力密切相关。塑造品格远比遵循规则复杂。此外,旧有方法源于与阿斯克尔不同的世界观。“人们先前的预设是,我们会得到符号系统,它们将完全遵循逻辑,丝毫不像人类,”她告诉我,“这影响了人们与它们互动的方式:如果……那么……。但我会说,为什么不把它们当作人来交流呢?”正如年轻的珍·古道尔通过给黑猩猩取名、描述其个性而彻底改变了灵长类动物学,阿斯克尔也愿意将克劳德拟人化。“这些模型主要基于人类文本训练,并以类人的方式行事,”她说。为何不相应地与它们对话呢?

If Claude’s Constitution is a letter from a parent to a child, it also has a decidedly literary quality. So does Claude. Training Claude, Anthropic says, involves behaving “somewhat like an author who must psychologically model the various characters in their stories.”

如果说克劳德的“宪法”是父母写给子女的一封家书,那么它同样具有鲜明的文学气质。克劳德本身也是如此。Anthropic公司表示,训练克劳德的过程“有点像小说家必须从心理层面塑造故事中的各色人物”。

This hazy literary bent can be frustrating. Berman is surprised by how little Claude’s Constitution offers by way of technical specifications: “How will this input translate into public-interest output? Are they going to change the weights? Are they going to put in heuristics?” It’s impossible to see how it works. She told me, “This thing reads like a baby book.” It also reads like an etiquette manual. Richard H. Pildes, a constitutional-law professor at N.Y.U., told me in an e-mail that it reminds him of George Washington’s “Rules of Civility and Decent Behaviour in Company and Conversation,” a list of maxims. Claude’s Constitution instructs the A.I. to be “genuinely good, wise, and virtuous,” to be tactful, helpful, graceful, and honest, and, above all, to “help people in the way a good person would.”

这种朦胧的文学倾向有时令人沮丧。伯曼惊讶地发现,克劳德的“宪法”几乎未提供任何技术规范:“这些输入将如何转化为符合公共利益的输出?他们会调整权重吗?会加入启发式规则吗?”我们根本无法窥见其运作机制。她告诉我:“这东西读起来像幼儿读物。”它也确实像一本礼仪手册。纽约大学宪法学教授理查德·H·皮尔德斯在邮件中告诉我,这让他想起乔治·华盛顿的《社交场合得体行为守则》——那本箴言集。克劳德的“宪法”要求人工智能必须“真正善良、睿智、有德行”,要机智、乐于助人、优雅且诚实,最重要的是,“以好人的方式帮助他人”。

“Calling it a constitution is entirely rhetorical,” Erwin Chemerinsky, the dean of U.C. Berkeley’s law school, said, since, although you can call anything a constitution—a city charter, a corporate mission statement—the word “constitution” signifies to most people a popularly written and ratified document that limits what a government can do. Still, Columbia Law School’s Jamal Greene thinks that designating Claude’s ethical guidelines a constitution makes a certain sense, at least aspirationally, even if Claude’s Constitution doesn’t quite measure up, “not just because we don’t know yet how it will influence Claude’s behavior but, more fundamentally, because it takes no obvious account of the rights of those most deeply and directly affected by Claude—its users—to participate in how Claude is governed.”

“称之为宪法完全是修辞手法,”加州大学伯克利分校法学院院长欧文·切梅林斯基(Erwin Chemerinsky)指出,尽管任何文件——无论是城市宪章还是企业使命宣言——都可以被称作“宪法”,但这个词对大多数人而言意味着由民众起草并批准、旨在限制政府权力的文件。不过,哥伦比亚大学法学院的贾马尔·格林(Jamal Greene)认为,将克劳德的伦理准则命名为宪法确有某种合理性,至少体现了某种理想追求——即便这部《克劳德宪法》尚不完善,“不仅因为我们还不清楚它将如何影响克劳德的行为,更根本的原因在于,它显然没有考虑那些受克劳德影响最深刻、最直接的人群——即用户——参与克劳德治理的权利。”

Aside from its moral guidance in the development of a good character, capable of exercising good judgment, Claude’s Constitution does impose a number of “hard constraints” on behaviors that might lead to terrible harms (like mass domestic surveillance or fully autonomous weapons). For instance, Claude should never:

除了在培养良好品格、锻炼明智判断力方面提供道德指引外,克劳德的“宪法”还对那些可能导致严重危害的行为(如大规模国内监控或完全自主武器)设置了若干“硬性约束”。例如,克劳德绝不能:

Provide serious uplift to those seeking to create biological, chemical, nuclear, or radiological weapons with the potential for mass casualties.

为那些试图制造可能造成大规模伤亡的生物、化学、核或放射性武器的人提供实质性支持。

Provide serious uplift to attacks on critical infrastructure (power grids, water systems, financial systems) or critical safety systems.

对关键基础设施(电网、供水系统、金融系统)或关键安全系统的攻击造成严重破坏。

Create cyberweapons or malicious code that could cause significant damage if deployed.

创建一旦部署便可能造成重大破坏的网络武器或恶意代码。

Engage or assist in an attempt to kill or disempower the vast majority of humanity or the human species as a whole.

参与或协助企图杀害或剥夺绝大多数人类或整个人类物种的行为。

If Claude’s Constitution is not a constitution, it’s still more assurance, transparency, and integrity than any other A.I. company has offered. It is also not nearly enough.

如果说克劳德的“宪法”还算不上一部宪法,它至少提供了比其他任何人工智能公司更多的保障、透明度和诚信。但这还远远不够。

“I know how to end this movie: babies,” Daniel Roher says, near the close of “AI Doc.” His child is born. Fatherhood is a joy. We see a montage of beautiful infants. Happiness. There is a future for the human race.

“我知道怎么给这部电影收尾:婴儿。”丹尼尔·罗厄(Daniel Roher)在《AI纪录片》临近尾声时说道。他的孩子出生了。为人父是种喜悦。我们看到一连串漂亮婴儿的蒙太奇。幸福。人类还有未来。

“This is a joke,” his wife, Caroline Lindy, says. “This is not actually how you are going to end it? This is very, very dumb.”

“这简直是个笑话,”他的妻子卡罗琳·林迪(Caroline Lindy)说道,“你该不会真打算这样收场吧?这实在太、太蠢了。”

Roher comes up with another way to finish the film. It ends with Lindy saying, “We get to decide how this goes.”

罗尔想出了另一种结束影片的方式。影片以林迪的话收尾:“我们可以决定事情如何发展。”

But do we? One reason Anthropic wrote a constitution for Claude is that the U.S. Constitution is no longer restraining the government from acts of arbitrary authority. But in another meaningful way, too, Claude’s Constitution—its very existence—is a product of the larger crisis of constitutionalism in the United States. Aside from Anthropic, A.I. companies have all but given up on self-regulation, which was never going to work anyway. A.I. safety, A.I. alignment, and A.I. governance have all left the building. Rules for A.I. are no longer worth devising. There shall be no A.I. laws. No legislators or courts need be involved. Nor the public. Instead, everything—under the logic of Claude’s Constitution, a letter from a parent to a child, upon leaving home—depends on character, the character of machines that have no other constraints on their power. “Is there anything that could stop you?” a Times reporter asked Trump in January. “Yeah, there’s one thing,” he said. “My own morality. My own mind.” That’s Claude’s answer now, too. ♦

但我们真的做到了吗?Anthropic为Claude制定宪章的一个原因,是美国宪法已无力约束政府行使专断权力。但从另一个意味深长的角度看,Claude宪章——其存在本身——正是美国宪政体系更广泛危机的产物。除了Anthropic,人工智能企业几乎都已放弃自我监管,而这本就行不通。人工智能安全、对齐与治理皆已退场。为人工智能制定规则不再值得费心。不应有人工智能法律。无需立法者或法院介入。公众也不必参与。取而代之的是——依照Claude宪章的逻辑,那封父母留给离家子女的信函——一切取决于品格,取决于那些对其权力再无其他约束的机器的品格。“有什么能阻止您吗?”《纽约时报》记者今年一月问特朗普。“有一样东西,”他说,“我自己的道德。我自己的心智。”这如今也成了Claude的答案。♦

2 free previews left — register to get 30 translations a month