Manuel Franco de la Peña challenges the conventional framing of model interpretability as a binary property. Rather than asking if a model is interpretable, the article argues we should focus on what specific questions explanations need to answer. Interpretability functions as a framework connecting humans and complex models, not as an inherent characteristic that can be checked off.

The author outlines three distinct functions of interpretability: diagnosing failures during development, validating learning in successful models, and extracting knowledge in domains where prediction alone is insufficient. These roles conceptually differ even when similar techniques are applied, requiring tailored approaches based on context and specific objectives.

Examples from medical imaging and computer vision demonstrate how interpretability methods reveal whether models learn meaningful patterns or superficial correlations. The article concludes that interpretability should be understood as a scientific tool rather than a vague requirement, with its value determined by the specific questions it enables us to answer.