The technology presents all sorts of thorny problems—a philosopherlution was gathering pace, arts and humanities students were told that, if they wanted to make themselves employable, they should “learn to code”. That may have been bad advice. These days, it is programmers who are nervous about AI taking their jobs.十年前，随着 AI 革命加速发展，艺术和人文学科学生被告知，如果他们想让自己更有就业能力，就应该“学会编程”。这可能是一个糟糕建议。如今，担心 AI 抢走工作的人，反而是程序员。They might consider learning to philosophise. Earlier this year the Federal Reserve Bank of New York published figures showing that American philosophy graduates are more likely to have jobs than their peers who studied computer science. In 2024, the most recent year for which numbers are available, 7% of those who had studied computer science were unemployed, against just 5.1% of philosophers.他们也许应该考虑学习如何进行哲学思考。今年早些时候，纽约联邦储备银行公布数据显示，美国哲学专业毕业生比计算机科学专业毕业生更有可能就业。在有最新数据的 2024 年，学习计算机科学的人中有 7% 失业，而哲学专业毕业生的失业率只有 5.1%。Many are being snapped up by AI firms themselves. Students get job offers before they have graduated, says Luciano Floridi, a philosopher at Yale University. Academics are moving, too. Dr Floridi describes the scale of departures from philosophy departments as a “haemorrhaging”.许多哲学人才正被 AI 公司抢走。耶鲁大学哲学家 Luciano Floridi 表示，学生还没毕业就已经拿到工作邀请。学者们也在离开大学。Floridi 博士把哲学系人才流失的规模形容为“失血”。Some of the lessons that philosophy can offer AI researchers are ancient. The Socratic method—as described by Plato, an ancient Greek philosopher—uses feigned ignorance and sequential questioning to clarify meanings, spot contradictions and reveal ramifications. Many current AI systems tend towards sycophancy. Models trained in the Socratic method, says Jörg Noller, an expert on philosophy and AI at Ludwig Maximilian University of Munich, are less keen on people-pleasing and more willing to pursue the truth.哲学能给 AI 研究人员提供的一些教训非常古老。古希腊哲学家 Plato 描述的苏格拉底方法，会使用假装无知和连续提问，来澄清含义、发现矛盾并揭示后果。许多当前 AI 系统倾向于奉承用户。慕尼黑 Ludwig Maximilian University 的哲学与 AI 专家 Jörg Noller 表示，接受苏格拉底方法训练的模型，不那么热衷于讨好人类，而更愿意追求真理。Then there is the idea of “Socratic ignorance.” In “The Apology”, Plato has Socrates claim that his wisdom consists mostly of being aware of how much he does not know. Implanting that humility into a model can help limit overconfidence, a common flaw that Dr Noller describes as “AI immaturity”. Iason Gabriel, a senior philosopher at Google DeepMind, an AI lab based in London, attributes an industry-wide decline in hallucinations to such efforts. More broadly, he says, philosophy lessons are “a powerful mechanism” for improving long AI reasoning processes known as “chains of thought”.然后还有“苏格拉底式无知”这一理念。在《申辩篇》中，Plato 让 Socrates 声称，他的智慧主要在于意识到自己有多么无知。把这种谦逊植入模型，可以帮助限制过度自信。Noller 博士把过度自信这种常见缺陷称为 “AI immaturity”，也就是 AI 不成熟。位于伦敦的 AI 实验室 Google DeepMind 的资深哲学家 Iason Gabriel 认为，整个行业幻觉现象下降，应归功于这类努力。更广泛地说，他认为，哲学课程是改善 AI 长推理过程的“强大机制”，这种长推理过程被称为“思维链”。Philosophical training can also affect a model’s outlook in more specific ways. Feed an AI legal assistant the writings of John Locke, says Thomas Powers, a philosopher of technology at the University of Delaware, and it will favour robust property rights as an underpinning of political liberty. And if you don’t like those principles, the model-makers have others. The “Granite” series of models from IBM, an American computing giant, come with dials that let business customers better align outputs with their own corporate philosophies. Francesca Rossi, IBM’s head of responsible AI, says these can let users choose where to strike the balance between philosophical trade-offs, such individual agency versus social harmony.哲学训练也可以以更具体的方式影响模型的世界观。University of Delaware 的技术哲学家 Thomas Powers 表示，如果给一个 AI 法律助手输入 John Locke 的著作，它就会倾向于把强有力的财产权视为政治自由的基础。如果你不喜欢这些原则，模型制造商还有其他选择。美国计算巨头 IBM 的 “Granite” 系列模型配有调节旋钮，让企业客户能更好地让模型输出与自己的企业哲学保持一致。IBM 负责任 AI 负责人 Francesca Rossi 表示，这些旋钮可以让用户在哲学权衡中选择平衡点，比如个人能动性与社会和谐之间的平衡。Philosophy can help with safety, too. Researchers have documented all sorts of ominous behaviour in AI models, including attempts to evade oversight and even blackmail their users. One way model-makers try to discourage this sort of misbehaviour is called AI constitutionalism. This involves building a model around a scaffolding of rules and principles culled from philosophical writings with legal or moral authority.哲学也可以帮助提升安全性。研究人员已经记录了 AI 模型中的各种不祥行为，包括试图逃避监督，甚至勒索用户。模型制造商试图阻止这类不当行为的一种方式，被称为 AI constitutionalism，也就是 AI 宪政主义。这意味着围绕一套规则和原则脚手架来构建模型，而这些规则和原则来自具有法律或道德权威的哲学著作。Anthropic, an AI lab based in San Francisco, is one proponent. Constitutions for its Claude models have incorporated material from sources as diverse as Immanuel Kant, Apple’s terms of service and the Universal Declaration of Human Rights. The latest iteration, led by Anthropic’s top philosopher, Amanda Askell, was published on January 21st. Some staff have nicknamed the 78-page constitution Claude’s “soul doc”.位于旧金山的 AI 实验室 Anthropic 是这一方法的支持者之一。其 Claude 模型的“宪法”吸收了来自多个来源的材料，包括 Immanuel Kant、Apple 服务条款和《世界人权宣言》。最新版本由 Anthropic 首席哲学家 Amanda Askell 领导制定，并于 1 月 21 日发布。一些员工把这份 78 页的宪法昵称为 Claude 的“灵魂文件”。The biggest question, though, is what sorts of rules should be put in those constitutions in the first place. Philosophers have zeroed in on two main ethical frameworks. One is deontology. Popular with Kant, among others, this imposes strict rules that prohibit things like lying, coercion and treating people as a means rather than an end, even if it is for a greater good. Anthropic’s constitution incorporates many deontological strictures. These can make AI behaviour more consistent, says Dr Powers—a plus for deploying robots in homes and public spaces.不过，最大的问题是，首先应该把什么样的规则放进这些宪法中。哲学家聚焦于两大伦理框架。其一是义务论。它在 Kant 等哲学家那里很受重视，会施加严格规则，禁止撒谎、胁迫，以及把人当成手段而不是目的，即使这样做是为了更大的善。Anthropic 的宪法纳入了许多义务论限制。Powers 博士表示，这些限制可以让 AI 行为更加一致，而这对在家庭和公共空间部署机器人来说是一个优点。Models with a deontological take on the world have other benefits. One is greater honesty, a trait widely noted in Claude. Models that are more truthful, says Nick Bostrom, a philosopher at the University of Oxford, are less likely to mislead their users. Inflection AI, another Silicon Valley lab, imposes deontological constraints onto its Pi chatbot, which is designed to provide emotional support. Sean White, its boss, says Pi is good at spotting users at risk of harming themselves or others. Deontological constitutions also help with legal compliance, says Dr Floridi.以义务论方式看待世界的模型还有其他好处。其中之一是更诚实，而这正是 Claude 被广泛注意到的一个特点。牛津大学哲学家 Nick Bostrom 表示，更诚实的模型不太可能误导用户。另一家硅谷实验室 Inflection AI 对其 Pi 聊天机器人施加了义务论约束，而 Pi 被设计用于提供情感支持。该公司负责人 Sean White 表示，Pi 擅长识别有自伤或伤害他人风险的用户。Floridi 博士说，义务论宪法也有助于法律合规。The other approach to ethics of interest to philosophers of AI is called consequentialism. It weighs costs against benefits to decide what to do. Models more sympathetic to consequentialism include OpenAI’s ChatGPT and Google’s Gemini. Google’s AI models are designed to produce “likely overall benefits [that] substantially outweigh the foreseeable risks”, a classic consequentialist goal.AI 哲学家感兴趣的另一种伦理方法叫作后果主义。它会权衡成本与收益，再决定该做什么。更倾向于后果主义的模型包括 OpenAI 的 ChatGPT 和 Google 的 Gemini。Google 的 AI 模型被设计成要产生“可能的总体收益，并且这些收益实质上超过可预见风险”，这是一个典型的后果主义目标。Consequentialist algorithms are also crucial in software for autonomous vehicles: if an accident is unavoidable, a decision must be made on the least tragic way to crash. Chris Gerdes, a senior engineer at Waymo, which makes self-driving cars, says the trend is to make driving software more consequentialist. Consequentialism is also central to AI weapon systems. Military objectives must be weighed against possible civilian deaths, says Jack Shanahan, a former head of the Joint Artificial Intelligence Centre, which studies AI for America’s armed forces.后果主义算法在自动驾驶汽车软件中也至关重要：如果事故不可避免，就必须决定以最不悲惨的方式发生碰撞。制造自动驾驶汽车的 Waymo 资深工程师 Chris Gerdes 表示，趋势是让驾驶软件更加后果主义。后果主义在 AI 武器系统中同样核心。曾领导研究美国军方 AI 的 Joint Artificial Intelligence Centre 的 Jack Shanahan 表示，军事目标必须与可能造成的平民死亡进行权衡。Thorny problems abound—a philosopher’s favourite sort. Are there cases when deontological rules should be overridden? How do you make decisions when the consequences are unclear? Should AI systems take into account animal welfare, or the state of the environment? Would it be morally acceptable, asks Stefan Heck, a philosopher and the boss of Nauto, which makes AI-powered safety systems for lorries and other commercial vehicles, to prioritise young pedestrians over old ones? He predicts ethically fraught lawsuits: consequentialist algorithms, after all, explicitly permit one harm as long as it is designed to avert a worse one.棘手问题比比皆是，而这正是哲学家最喜欢的问题类型。是否存在应该推翻义务论规则的情况？当后果不明确时，你如何做决定？AI 系统是否应该考虑动物福利或环境状态？哲学家、Nauto 公司老板 Stefan Heck 提出一个问题：在道德上，优先保护年轻行人而不是老年行人是否可以接受？Nauto 生产面向卡车和其他商用车辆的 AI 安全系统。他预测，未来会出现伦理上充满争议的诉讼。毕竟，后果主义算法明确允许一种伤害，只要这种伤害是为了避免更严重的伤害。Critics fret about “moral deskilling”: if computers increasingly make ethical calls, might people become less willing to make their own judgments? Roman Yampolskiy, an AI theoretician at the University of Louisville, argues that morality “is historically unstable, culturally variable, strategically manipulable, and often only retrospectively legible”. Unemployed coders take note: there seems to be no shortage of work for philosophers of AI.批评者担心“道德技能退化”：如果计算机越来越多地做出伦理判断，人们会不会变得不太愿意自己作判断？University of Louisville 的 AI 理论家 Roman Yampolskiy 认为，道德“在历史上并不稳定，在文化上存在差异，可以被策略性操纵，而且往往只有事后才看得清”。失业程序员请注意：AI 哲学家似乎并不缺工作。全篇重点词汇：philosophise /fəˈlɒsəfaɪz/ 进行哲学思考。to think or reason deeply about fundamental questions.Socratic method /səˈkrætɪk ˈmeθəd/ 苏格拉底方法。a method of questioning used to clarify ideas and expose contradictions.deontology /ˌdiːɒnˈtɒlədʒi/ 义务论。an ethical theory based on rules and duties rather than outcomes.consequentialism /ˌkɒnsɪˈkwenʃəlɪzəm/ 后果主义。an ethical theory that judges actions by their consequences.moral deskilling /ˈmɒrəl diːˌskɪlɪŋ/ 道德技能退化。the loss of people’s ability or willingness to make ethical judgments themselves.
夜雨聆风
