制度回放：六起 AI 失败的防线反事实检验
Replay as Governance: Counterfactual Barrier Analysis of Six AI Failures
摘要
AI 治理文献正在产出越来越多的协议、框架与原则，但没有一篇验证过这些制度在真实事故中能否拦住失败。本文提出一种制度回放方法（Barrier Replay Protocol, BRP）：选取已公开的 AI 失败事件，还原其制度防线栈（每层标「在场且有效、在场但失效、不在场」三态），画出失效链，然后在时间线上注入我们的制度构件（闸门、错误记忆、任期制、审计、分层防御），回答一个受限的反事实问题：若该构件在场，能否在暴露点之前拦截。我们对六起跨域跨时代的事件做了回放：Air Canada 聊天机器人赔偿案、Mata v. Avianca 律师假引用案、Tay 种族歧视事件、COMPAS 风险评估偏差、波音 737 MAX MCAS 空难、武汉长江偷倒黑泥污染事件。回放显示：六起事件的 30 道理论防线中有 26 道失效或不在场；出站闸门参与 4 起事件的介入方案，其中 1 起单靠闸门即可拦截、5 起需双构件协同；最具横向可迁移性的构件是出站闸门与跨部门数据打通（数据孤岛型失效的正面解法）。本文同时把六起事件的构件映射到 EU AI Act 第 14.4/14.5 条与中国《生成式人工智能服务管理暂行办法》《人工智能生成合成内容标识办法》的具体条文，给出一张「制度构件与法规条文对照表」。回放的输出不是事后评论，而是可检验的制度假设：下一个 AI 失败事件可以用来证伪或加强本文的拦截判定。
Abstract
AI governance literature is producing more and more protocols, frameworks and principles, yet none has verified whether these institutions could stop failure in real incidents. This paper proposes a method of institutional replay (Barrier Replay Protocol, BRP): take publicly documented AI failure events, reconstruct their institutional barrier stack (each layer marked as present-and-effective, present-but-failed, or absent), draw the failure chain, then inject our institutional components (gates, error memory, tenure, audit, layered defense) on the timeline to answer a bounded counterfactual question: had the component been present, could it have intercepted before the exposure point. We replayed six events across domains and eras: the Air Canada chatbot compensation case, the Mata v. Avianca fake-citation case, the Tay racist-content incident, COMPAS risk-assessment bias, the Boeing 737 MAX MCAS crashes, and the Wuhan Yangtze illegal sludge-dumping pollution incident. The replay shows that 26 of the 30 theoretical barriers across the six events failed or were absent; the outbound gate enters the intervention plan of four events, in one of which the gate alone suffices while five require two components; the most horizontally transferable components are the outbound gate and cross-department data connection (a direct remedy for data-silo-type failures). We also map the components of the six events onto Articles 14.4/14.5 of the EU AI Act and specific provisions of Chinas Interim Measures for the Management of Generative AI Services and the Measures for Labeling AI-Generated Synthetic Content, and provide a table mapping institutional components to regulatory provisions. The output of replay is not retrospective commentary but testable institutional hypotheses: the next AI failure event can falsify or strengthen our interception judgments.关键词
制度回放/反事实分析/AI 治理/防线/闸门/分层防御/AI 事故Key words
institutional replay/counterfactual analysis/AI governance/barriers/gates/layered defense/AI incidents引用本文复制引用
臧存勋.制度回放：六起 AI 失败的防线反事实检验[EB/OL].(2026-10-29)[2026-11-02].https://chinaxiv.org/abs/202610.00234.学科分类
计算技术、计算机技术
