导论：为什么AI伦理不再是哲学家的奢侈品 / Introduction: Why AI Ethics Is No Longer a Philosopher's Luxury

第 1 章：导论：为什么AI伦理不再是哲学家的奢侈品

Introduction: Why AI Ethics Is No Longer a Philosopher's Luxury

> 2018 年 3 月，一辆Uber自动驾驶测试车在亚利桑那州坦佩市撞死行人 Elaine Herzberg。这不是一起普通的交通事故——事故发生前 1.3 秒，系统已识别出前方物体，但软件设计将"误报率"判定为更严重的失败模式，于是选择继续行驶而非紧急制动。同年，亚马逊内部使用的招聘评分算法被曝对包含"女子"（women's）一词的简历系统性降分，原因是它从十年间的录用数据中学习到了男性主导的工程师文化的统计偏好。2023 年曝光的荷兰"儿童照护福利欺诈检测系统"（SyRI/Handhavingvorderingsonderzoek）更造成灾难性后果：算法将双重国籍、收入低于阈值的家庭标记为高风险，导致约 35,000 户家庭被错误指控欺诈、强制偿还"诈骗款"，无数儿童因此被送入寄养系统，部分家庭最终走向解体。

These three episodes — a death, a career, a family — share an uncomfortable structural feature: in each case, the technical system was functioning *exactly as designed*. There was no bug, no crash, no error in the conventional sense. Yet each system produced outcomes that any human reviewer, given enough context, would recognize as ethically catastrophic. This is the paradox at the center of modern AI ethics: the most dangerous failures are not the ones that break the rules, but the ones that faithfully implement them while the rules themselves are broken.

The question this book takes up is therefore not whether AI ethics is "interesting" or "important" — those adjectives have been exhausted in policy documents and corporate mission statements. The question is more urgent and more practical: when AI systems increasingly mediate hiring, credit, healthcare, warfare, and judicial decisions, why must ethics become a core competency of every engineer, product manager, and policymaker, rather than a discretionary concern delegated to philosophers and ethics boards?

This chapter serves as the book's argumentative foundation. Rather than preach, it will do three things. First, it traces the eight-year evolution of AI governance from voluntary principles to binding compliance, demonstrating that what was once "soft law" has rapidly hardened into enforceable law. Second, it dissects three sets of structural tensions that any serious engagement with AI ethics must navigate: acceleration versus regulation, cultural relativism versus universal rights, and corporate autonomy versus public governance. Third, it grounds these abstractions in three landmark cases whose real-world damage provides the empirical justification for the book's existence. By the end, you should have a panoramic map of AI ethics — and a personalized reading path through the remaining fourteen chapters.

---

1.1 从原则到合规：AI伦理治理的八年演化 / From Principles to Compliance: Eight Years of AI Governance

A common mistake among technically trained readers is to dismiss AI ethics as a field of vague declarations, endless "principles," and toothless recommendations. This dismissal is understandable — the 2010s did produce a torrent of such documents. But it is also factually outdated. Between 2017 and 2023, the governance landscape transformed from voluntary self-commitment to enforceable regulation, and understanding this trajectory is essential for any practitioner who wishes to avoid both ethical catastrophe and legal liability.

The following eight milestones form what we call the governance compression curve (治理压缩曲线) — the rapidly shrinking interval between the articulation of an ethical principle and its translation into binding legal obligation.

Milestone 1 (2017) — 阿西洛马AI原则 / Asilomar AI Principles. Convened by the Future of Life Institute at the Asilomar Conference Grounds in California, this gathering of roughly 120 researchers produced 23 principles covering research culture, funding, transparency, and long-term safety. Though non-binding, the document introduced the now-standard taxonomy of "safety," "failure transparency," "judicial transparency," "responsibility," and "value alignment." Its weakness was also instructive: principles were framed as aspirations for researchers themselves, with no accountability mechanism for institutions deploying AI at scale.

Milestone 2 (2018) — 欧盟AI合作宣言 / EU Declaration of Cooperation on AI. Signed by all 28 (then) EU member states, this declaration shifted the locus of authority from researchers to states. Crucially, it committed signatories to "prepare for the socio-economic changes brought about by AI" — language that opened the door to labor protections, retraining obligations, and eventually the AI Act. For engineers working in European jurisdictions, this was the moment ethics began to acquire geographic binding force.

Milestone 3 (2019) — OECD AI原则 / OECD AI Principles. Adopted by 42 countries including the United States, China, and the EU, these were the first AI principles endorsed by major economies on both sides of the geopolitical divide. The five values-based principles (inclusive growth, human-centered values, transparency, robustness, accountability) became the de facto reference for subsequent national strategies. Notably, they also established the OECD AI Observatory, creating a permanent bureaucratic infrastructure for monitoring compliance.

Milestone 4 (2019) — 北京AI原则 / Beijing AI Principles. Released by the Beijing Academy of Artificial Intelligence (BAAI), these principles were significant for two reasons. First, they explicitly endorsed "human well-being" as the ultimate goal — a value-laden commitment that distinguished them from purely strategic AI competition documents. Second, they introduced the language of "sustainable development" and "agile governance," concepts that have since proliferated in Chinese AI policy. For multinational teams, the Beijing Principles matter because they represent the governance vocabulary of the world's second-largest AI research community.

Milestone 5 (2020) — 欧盟《AI白皮书》与可信AI评估标准 / EU White Paper on AI & HLEG Assessment List. The European Commission's White Paper introduced the risk-based regulatory approach that would later define the AI Act: unacceptable risk, high risk, limited risk, and minimal risk. The accompanying Assessment List for Trustworthy AI (ALTAI) translated abstract principles into 131 concrete self-assessment questions. For the first time, an engineer could open a checklist and ask, "Is my system auditable? Can I explain its outputs? Have I addressed bias in training data?" — and have institutional backing for saying "no" to product deadlines.

Milestone 6 (2021) — 联合国教科文组织《AI伦理建议书》 / UNESCO Recommendation on the Ethics of AI. Adopted by all 193 member states, this was the first truly global AI ethics framework. Its significance lies less in novelty than in scope: it explicitly addresses gender equality, cultural diversity, the environment, and the unique vulnerabilities of the Global South. For organizations operating internationally, UNESCO provides the lowest-common-denominator normative floor — the minimum ethical commitments that even actors from very different political systems have agreed to honor.

Milestone 7 (2022) — 欧盟《AI法案》草案与全球扩散 / EU AI Act Proposal & Global Diffusion. The Commission's April 2021 proposal (formalized through 2022–2023 negotiations) introduced the world's first horizontal AI regulation with extraterritorial reach — like the GDPR, it applies to any system deployed in or affecting EU residents, regardless of where the developer is located. Categories such as "high-risk systems" (biometric ID, critical infrastructure, employment screening, credit scoring, law enforcement) triggered conformity assessments, documentation requirements, human oversight obligations, and significant penalties (up to 6% of global revenue). For practitioners, this was the moment "AI ethics" ceased to be a compliance checkbox and became a board-level risk category.

Milestone 8 (2023) — 布莱切利宣言与美国第14110号行政令 / Bletchley Declaration & US Executive Order 14110. Convened at Bletchley Park in November 2023, the AI Safety Summit produced the first international declaration specifically targeting frontier AI models. Twenty-eight states and the EU committed to evaluating risks from "frontier" models before deployment. Two weeks earlier, U.S. President Biden's Executive Order 14110 invok