Scholarlink

<p>As more complex Artificial Intelligence Systems (AISs) become increasingly embedded in critical sectors, the “black box dilemma” has emerged as a key concern in both technical and legal debates. Unfortunately, the concepts most frequently invoked to formulate and address the epistemic dimension of AISs—<i>transparency</i>, <i>traceability</i>, <i>interpretability</i>, and <i>explainability</i>—remain ill-defined and inconsistently applied in EU regulation. This article proposes a regulatory taxonomy that distinguishes these four concepts as layered and interdependent dimensions of AI <i>opacity</i>, each with distinct epistemic and normative roles. While each of these concepts offers a necessary but partial view of AI opacity, none is sufficient on its own. They support a complete understanding of AIS outcomes only when considered together, as interdependent but connected layers. Thus, the taxonomy provides a conceptual framework for legal interpretation, compliance strategies, and informed future legislative design. The article illustrates the new framework through a case study on algorithmic credit scoring. By clarifying the distinct functions and audiences of each concept, the article contributes to a more coherent regulatory approach to AI opacity, one that enables accountability, fosters innovation, and strengthens trust in automated decision-making.</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

A regulatory taxonomy of AI opacity in the EU: rethinking transparency, traceability, interpretability, and explainability

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Carlotta Buttaboni,
  • Luciano Floridi

摘要

As more complex Artificial Intelligence Systems (AISs) become increasingly embedded in critical sectors, the “black box dilemma” has emerged as a key concern in both technical and legal debates. Unfortunately, the concepts most frequently invoked to formulate and address the epistemic dimension of AISs—transparency, traceability, interpretability, and explainability—remain ill-defined and inconsistently applied in EU regulation. This article proposes a regulatory taxonomy that distinguishes these four concepts as layered and interdependent dimensions of AI opacity, each with distinct epistemic and normative roles. While each of these concepts offers a necessary but partial view of AI opacity, none is sufficient on its own. They support a complete understanding of AIS outcomes only when considered together, as interdependent but connected layers. Thus, the taxonomy provides a conceptual framework for legal interpretation, compliance strategies, and informed future legislative design. The article illustrates the new framework through a case study on algorithmic credit scoring. By clarifying the distinct functions and audiences of each concept, the article contributes to a more coherent regulatory approach to AI opacity, one that enables accountability, fosters innovation, and strengthens trust in automated decision-making.