Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead

Cynthia Rudin is a professor of computer science, electrical and computer engineering, and statistical science at Duke University, where she runs the Interpretable Machine Learning Lab. Before this paper she had built inherently transparent models for consequential settings — CORELS and related certifiably optimal rule lists, sparse scoring systems for medical and criminal-justice use — so the position the paper argues is the one her research programme already embodied. The work was recognised when she received the 2022 AAAI Squirrel AI Award for AI for the benefit of humanity, an award tied closely to this line of argument.

The paper appeared in 2019 in Nature Machine Intelligence, then a newly launched member of the Nature family, as a peer-reviewed Perspective rather than an experimental report. Its argument is built on a distinction it presses hard: an explanation is a separate, simpler model that approximates a black box's outputs after the fact and can disagree with the thing it explains, whereas an interpretable model is constrained to be understandable from the start. Rudin's case is that in high-stakes domains — recidivism prediction, medical decisions, parole — the accuracy lost by demanding interpretability is often slight or absent, and the COMPAS recidivism tool serves as her running example of opacity protecting a system from scrutiny.

The piece sits as a position paper, secondary to the primary methods it criticises and to her own technical work, which it cites and is meant to generalise. It states the "do not explain, replace" pole of the interpretability debate, and it reads directly against Ribeiro, Singh and Guestrin's LIME, whose promise was to explain any classifier post hoc. It shares its suspicion of after-the-fact explanation with Lipton's Mythos of Model Interpretability, though Lipton deflates the very vocabulary of "interpretable" that Rudin builds on, so the two skepticisms point in different directions.

Rudin's stake is professional and ideological in the ordinary academic sense. The paper is a manifesto for the research agenda she is identified with, and its force — that explaining black boxes is not merely second-best but harmful — raises the standing of the interpretable-models work she and her students produce. There is no commercial conflict; the interest is reputational, and the paper wears its advocacy openly rather than presenting itself as a neutral survey.