ML Interpretability: Simple Isn't Easy

Interpretability is not a single concept: The paper demonstrates that "simple" models like linear models and decision trees achieve interpretability in different ways, suggesting that interpretability itself is heterogeneous rather than a one-size-fits-all property that can be uniformly applied across all AI systems.

"Simple" models require careful design to remain interpretable: Despite their reputation for transparency, even straightforward models like linear regression and decision trees require deliberate choices about features and structure to maintain interpretability—simplicity alone does not guarantee understanding.

Mid-complexity models can preserve some interpretability: The paper examines MARS (Multivariate Adaptive Regression Splines) and GAM (Generalized Additive Models) as examples of more complex approaches that retain degrees of interpretability, suggesting a spectrum rather than a binary choice between "black box" and "transparent."

Clarity about what interpretability means matters for responsible AI: By clarifying what interpretability actually is "in particular cases," the paper suggests that responsible AI governance requires precise definitions tailored to specific contexts rather than vague appeals to transparency or explainability.

The focus on black-box models may obscure important questions: By examining why simpler models work, the paper redirects attention from just fixing opaque neural networks to understanding the foundational principles of interpretability itself.