What Kind of Claims Are Transparency, Explainability, and Interpretability? A Definitional Taxonomy for Health AI.

Transparency, explainability, and interpretability are ubiquitous in the health artificial intelligence (health AI) literature, yet are used inconsistently and often interchangeably. This Viewpoint does not attempt to define the three terms. It asks a prior, meta-level question: what kind of claim do we make when we call a system transparent, explainable, or interpretable? I argue that the three terms pick out three different kinds of claim. A transparency claim is structural: it concerns what can be inspected about a system. An explainability claim is relational: it concerns whether understanding is successfully conveyed to a specified class of observer. An interpretability claim is functional: it concerns whether a bounded human agent can, in principle, re-derive the system's output by following its procedure. I use the framing example of a clinical decision support system that uses machine learning to identify patients at risk of deterioration. I develop each claim type through competing technical accounts, consider salient objections to each, and defend the taxonomy against them. I close by drawing out what the taxonomy implies for the design, evaluation, reporting, and regulation of health AI, and for physician autonomy at the point of care.