Scholarlink

<p>This paper introduces the “Interpreter’s Trap,” a socio-legal and STS-informed (Here, “STS-informed” refers to a Science and Technology Studies orientation that treats algorithmic categories (e.g., “risk”) and their evidentiary status as sociotechnical accomplishments shaped by institutional practices, rather than as neutral representations of a stable ground truth.) framework developed through a concept-generative case study of the COMPAS recidivism risk assessment tool. Moving beyond individual-level cognitive-deficit accounts of reliance, I argue that the trap is best understood as an epistemic double bind that reallocates justificatory burdens under constrained contestability and asymmetric liability—even when decision-makers retain nominal discretion and may deviate from the algorithm’s outputs. Interpreters are positioned between a “contaminated objectivity” rooted in proxy-laden predictions and disputed ground truth, and an “eroded subjectivity,” where exercising discretion becomes institutionally costlier by concentrating the&#xa0;de facto institutional burden of justification and potential blame on the individual judge. The analysis shows that explainable AI (XAI) often provides technical rationales without supplying normative reasons that are legally reviewable, instead functioning as a catalyst for “accountability washing” by offering convenient narratives that mask the system’s “complexity illusion.” In the age of Generative AI, language-model explanation layers may intensify the trap by producing persuasive yet unfaithful rationales. I therefore advocate a paradigm shift from explaining black boxes to designing inherently interpretable “glass-box” models—particularly for structured-data, high-stakes normative decisions—as a necessary (though not sufficient) precondition to realize the legal “Right to Contest” and enable meaningful Human Oversight, paired with contestation infrastructures. This paper contributes (1) a meso-level mechanism of justificatory risk allocation—under constrained contestability and asymmetric liability—showing how post hoc XAI supplies defensibility more than understanding; and (2) governance implications: why meaningful Human Oversight in high-stakes normative decisions requires design-based interpretability and contestation infrastructure.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The interpreter’s trap: how explainable AI launders uncertainty into justification—a socio-legal case study of COMPAS risk assessment

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Keita Sode

摘要

This paper introduces the “Interpreter’s Trap,” a socio-legal and STS-informed (Here, “STS-informed” refers to a Science and Technology Studies orientation that treats algorithmic categories (e.g., “risk”) and their evidentiary status as sociotechnical accomplishments shaped by institutional practices, rather than as neutral representations of a stable ground truth.) framework developed through a concept-generative case study of the COMPAS recidivism risk assessment tool. Moving beyond individual-level cognitive-deficit accounts of reliance, I argue that the trap is best understood as an epistemic double bind that reallocates justificatory burdens under constrained contestability and asymmetric liability—even when decision-makers retain nominal discretion and may deviate from the algorithm’s outputs. Interpreters are positioned between a “contaminated objectivity” rooted in proxy-laden predictions and disputed ground truth, and an “eroded subjectivity,” where exercising discretion becomes institutionally costlier by concentrating the de facto institutional burden of justification and potential blame on the individual judge. The analysis shows that explainable AI (XAI) often provides technical rationales without supplying normative reasons that are legally reviewable, instead functioning as a catalyst for “accountability washing” by offering convenient narratives that mask the system’s “complexity illusion.” In the age of Generative AI, language-model explanation layers may intensify the trap by producing persuasive yet unfaithful rationales. I therefore advocate a paradigm shift from explaining black boxes to designing inherently interpretable “glass-box” models—particularly for structured-data, high-stakes normative decisions—as a necessary (though not sufficient) precondition to realize the legal “Right to Contest” and enable meaningful Human Oversight, paired with contestation infrastructures. This paper contributes (1) a meso-level mechanism of justificatory risk allocation—under constrained contestability and asymmetric liability—showing how post hoc XAI supplies defensibility more than understanding; and (2) governance implications: why meaningful Human Oversight in high-stakes normative decisions requires design-based interpretability and contestation infrastructure.