What Does Explainable AI Really Mean? A New Conceptualization of Perspectives

Explainable AI means different things across fields: The paper's analysis of conference proceedings from computer vision, natural language processing, cognitive science, and machine learning reveals that researchers use "explainable AI" inconsistently, with different communities prioritizing different approaches to making AI systems understandable.

There are three established levels of AI transparency: Current systems fall into three categories—opaque (no insight into how decisions are made), interpretable (users can mathematically analyze the mechanisms), and comprehensible (systems provide symbols that help users understand conclusions)—each offering different degrees of transparency.

Most AI explanations still require human interpretation: The paper identifies a significant gap in current practice: most "explainable" systems generate outputs that need human post-processing to become truly understandable, meaning the explanation process isn't fully automated and remains dependent on human expertise.

A new standard for true explainability is needed: The authors propose a fourth category—"truly explainable systems"—where the AI itself generates complete, human-ready explanations through automated reasoning, without requiring additional human work to make sense of the output.

Clarifying terminology matters for responsible AI: By distinguishing between these different notions of explainability, the paper highlights that claims about an AI system being "explainable" may not mean what stakeholders, regulators, or users assume they mean.