在鼓励人文学科学生“学习写代码”的十年后，就业市场的转变使程序员对人工智能的排挤感到焦虑，而哲学毕业生却找到了强劲的就业机会。根据纽约联邦储备银行2024年的数据，计算机科学毕业生的失业率为7%，而哲学毕业生仅为5.1%。人工智能公司正在积极招聘哲学家，在他们毕业前就提供工作机会，并促使学术界人士从大学系所大量流失。这些专业人员应用哲学理论来解决人工智能的关键挑战，例如使用苏格拉底方法减少模型的阿谀奉承，以及通过“苏格拉底式的无知”来遏制幻觉和过度自信。

哲学框架还有助于定制模型行为并提高安全性。例如，IBM的Granite模型提供了可定制的设置，以使输出与特定的公司哲学保持一致。为了强制执行安全性并防止用户遭受勒索或规避监管，实验室利用“人工智能宪政主义”——围绕法律和道德规则构建模型。Anthropic的Claude使用了一份78页的宪法，该宪法借鉴了康德、苹果的服务条款和人权宣言。义务论框架禁止撒谎和强迫（无论结果如何），这在Anthropic的Claude和Inflection AI's Pi中很受欢迎，促进了诚实、法律合规以及更安全的物理和情感互动。

相反，后果论（通过权衡成本与收益以最大化整体利益）指导着像OpenAI的ChatGPT和Google的Gemini这样的模型。后果论算法对于决定如何最小化碰撞影响的自动驾驶汽车，以及权衡军事目标与平民风险的军事人工智能系统至关重要。然而，这些系统引入了复杂的道德困境，例如是否优先保护年轻行人而非年长行人，这可能会引发极具挑战性的诉讼。批评者还警告说，随着人类将道德决策委托给算法，可能会出现“道德去技能化”，这表明对人工智能哲学家导航这些不稳定且随文化变化的道德领域有着持续增长的长期需求。

A decade after humanities students were urged to "learn to code," a shift in the job market has left programmers anxious about AI displacement while philosophy graduates find robust opportunities. According to 2024 Federal Reserve Bank of New York data, computer science graduates faced a 7% unemployment rate, compared to just 5.1% for philosophy graduates. AI firms are aggressively recruiting philosophers, drawing students before graduation and prompting a significant exodus of academics from university departments. These professionals apply philosophical theories to address key AI challenges, such as reducing model sycophancy using the Socratic method and curbing hallucinations and overconfidence through "Socratic ignorance."

Philosophical frameworks also help customize model behaviors and enhance safety. For example, IBM's Granite models offer customizable settings to align outputs with specific corporate philosophies. To enforce safety and prevent user blackmail or evasion of oversight, labs utilize "AI constitutionalism"—structuring models around legal and moral rules. Anthropic's Claude uses a 78-page constitution drawing from Kant, Apple's terms, and human rights declarations. Deontological frameworks, which prohibit lying and coercion regardless of outcomes, are popular in Anthropic's Claude and Inflection AI's Pi, promoting honesty, legal compliance, and safer physical and emotional interactions.

Conversely, consequentialism, which weighs costs against benefits to maximize overall good, guides models like OpenAI's ChatGPT and Google's Gemini. Consequentialist algorithms are critical for autonomous vehicles deciding how to minimize collision impact, and for military AI systems balancing objectives against civilian risks. However, these systems introduce complex ethical dilemmas, such as whether to prioritize saving younger pedestrians over older ones, which could spark challenging lawsuits. Critics also warn of "moral deskilling" as humans delegate ethical decisions to algorithms, indicating a growing, long-term demand for AI philosophers to navigate these unstable and culturally variable moral landscapes.

Source: Why big AI labs are hiring so many philosophers

Subtitle: The technology presents all sorts of thorny problems—a philosopher’s favourite kind

Dateline: 6月 25, 2026 04:19 上午