Review Interpretability

Interpretability and explainability describe methods and system qualities that help people understand how a model works, which factors influenced a result, or why a particular output was produced.

Interpretability often concerns how understandable the model itself is. Explainability often concerns an explanation supplied for behavior or a specific result. Usage varies across fields and regulations.

People may need to diagnose errors, contest decisions, meet governance obligations, or decide whether a result is appropriate for the real situation.

A lending system shows the factors that materially affected a decision and gives reviewers enough information to test whether the explanation matches actual model behavior.

An explanation is not automatically faithful, sufficient, or proof that the decision was correct or fair.

Make behavior understandable enough to inspect and challenge.

* [[verification|Verification]] * [[algorithmic-accountability|Algorithmic Accountability]] * [[model-card|Model Card]] * [[awareness-boundary|Awareness Boundary]] * [[chain-of-thought|Chain of Thought]]

* [NIST, *Four Principles of Explainable Artificial Intelligence* (NISTIR 8312)](https://doi.org/10.6028/NIST.IR.8312) — principles for explanations, meaningfulness, explanation accuracy, and knowledge limits. * [NIST, *Psychological Foundations of Explainability and Interpretability in Artificial Intelligence* (NISTIR 8367)](https://doi.org/10.6028/NIST.IR.8367) — examines distinctions between explainability and interpretability and their human context.

* Distinguish explanation from evidence through [[verification|Verification]].