Explainable AI vs interpretable AI

Explainable AI (XAI) is gaining a lot of attention these day, and there is a lot of talk about model interpretability. But whare are the similarities and differences between these terms?

Well, for starters, the terms interpretability and explainability are used interchangeably in the literature (Adadi & Berrada, 2018). According to Lewis (1986, p. 217), to explain an event means to provide information about its causal history, and an explanation can thus be considered an answer to a why-question (Miller, 2018). The focus on causes is echoed in Kim, Khanna & Koyejo (2016) who define interpretability as the degree to which a human can consistently predict the model’s result. Several researchers consider explainability and interpretability to be two sides of the same coin (Miller, 2018; Lipton, 2017; Molnar, 2021), and explainability is sometimes construed as post-hoc interpretability (Miller, 2018; Lipton, 2017). Lipton (2017), Molnar (2021), Rudin (2019) and Adadi & Berrada (2018) all consider interpretable models to include rule based systems, decision trees and logistic and linear regression models and their extensions (e.g. GAM) which are constrained in model form, e.g. by obeying monotonicity or additivity. It should be noted, though, that very large decision trees and linear models with many features can become uninterpretable (Lipton, 2017).

Techniques and model properties that either enable or comprise interpretability generally fall into two categories: 1) transparency, i.e. understanding how the model works, 2)…