Finale Doshi-Velez and Been Kim, 2017

“Interpretability is the ability to explain or to present in understandable terms to a human.”

Finale Doshi-Velez and Been Kim wrote in 2017 that machine-learning interpretability needed a more rigorous science. Explanations were becoming common in safety, health, and fairness work, yet researchers often used the word “interpretable” without agreeing on what it meant or how to evaluate it. Their definition placed a human recipient at the center and connected explanation quality to the task that person must perform. The wording is compact because it identifies the change the authors wanted readers to notice, rather than turning a complex project into a promise of automatic progress.

Understandability is not a property of an artifact in isolation. A display that helps a data scientist debug a model may not help a patient decide whether to trust a recommendation. The authors distinguished application-grounded evaluation, human-grounded studies, and functionally grounded proxies. That taxonomy encourages researchers to select evidence proportionate to the real context rather than assuming that sparsity, a visualization, or a short explanation is universally meaningful. A careful reading separates the mechanism from the mythology that later accumulates around it. The quotation becomes more useful when its assumptions remain visible.

Interpretability claims now appear in commercial platforms, regulations, and model documentation. Teams should begin by identifying the audience and decision: debugging, recourse, scientific discovery, oversight, or informed consent. They can then test whether the explanation improves accuracy, speed, error discovery, or confidence calibration for those users. A polished interface is not sufficient evidence if people cannot act more effectively because of it. The modern relevance lies in method, not nostalgia. Researchers and product teams can preserve the original curiosity while demanding stronger documentation, independent checks, and realistic operating conditions.

Explanations can persuade even when they are incomplete or unfaithful. They may shift responsibility onto users who lack the authority or time to challenge a system. Rigorous evaluation should include misleading cases, uncertainty, and the costs of misunderstanding. Sometimes the better response is to simplify the model or redesign the decision process rather than adding another explanation layer. A responsible interpretation keeps people in the frame. Technical success is incomplete if the system weakens agency, obscures accountability, or shifts risk onto those with the least power to object. The definition’s simplicity is demanding: an explanation succeeds only in relation to a human purpose. Interpretability should be tested as carefully as predictive performance. For readers, the practical habit is to return to the evidence: identify the task, examine the conditions of success, and ask who can act when reality differs from the demonstration. The work continues with care.