interpretables

Interpretables refer to models, explanations, or representations in artificial intelligence that humans can understand. Interpretability measures how well a system's decisions and the factors driving them can be traced. The aim is transparency, trust, and accountability, especially in domains with significant consequences. Interpretables are contrasted with opaque black-box models whose inner workings are hard to inspect.

Approaches fall into intrinsic interpretability, using transparent models such as linear models, decision trees, or rule-based systems; and post-hoc interpretability, which explains a trained model using techniques like feature attribution (SHAP, LIME), surrogate models, or counterfactuals. Explanations can be global (model-wide) or local (per prediction).

Applications include healthcare, finance, law, and public policy, where explanations support decision justification and auditing. Evaluation blends technical criteria—fidelity, simplicity, stability—with human-centered assessments of usefulness and trust. A common trade-off is between interpretability and accuracy, and explanations may introduce biases or mislead if not designed carefully.

Challenges include definitional ambiguity, domain-specific notions of understandability, and ensuring explanations reflect real causes rather than correlations. Ongoing research addresses better evaluation methods, causal explanations, fairness, and robustness, as well as standards for reporting what explanations cover.