The Accuracy-Explainability Trade-off, the Right to Explanation, and the Implications for Organisations
研究了在算法决策中，解释权所要求的可靠性和可验证性与准确性之间的冲突，认为自我辩护利益不足以支撑普遍的解释权，组织应基于合法性、公平性和信任来履行解释义务。
Abstract Algorithms increasingly shape access to employment, credit, healthcare, and justice, yet the basis on which they do so is often opaque. A growing literature argues that affected individuals have a right to explanation, grounded in their interest in informed self-advocacy: the ability to understand and respond to decisions that bear on their life prospects. We examine whether this interest can sustain such a right. Explanations that serve self-advocacy must be reliable (truth-tracking) and verifiable (open to independent check). We show that in open-ended decision environments, where evaluative criteria must be discovered rather than stipulated in advance, reliability and verifiability conflict with accuracy. This trade-off arises in human decision-making and has a structural analogue in AI systems. Because requiring thick explanation would systematically reduce the quality of decisions, the self-advocacy interest cannot by itself ground a general right. Where thick explanation is nonetheless owed, it is justified by different grounds: legality, fairness, and trust. We draw out implications for how organisations should structure their obligations.