Explainable AI

Different audiences need different explanations. A developer may need to diagnose a failure, an operator may need to decide when to defer and an affected person may need to understand or contest a decision. Global explanations describe broad behavior; local explanations describe a particular output. Feature attributions, examples and counterfactuals answer different questions and depend on assumptions about the data and model. For generative AI, a fluent rationale is especially easy to mistake for evidence of internal reasoning. Explanation quality therefore includes fidelity to the relevant process, usefulness to the audience and explicit recognition of what remains unknown.

A practitioner chooses an explanation method from the decision it must support and verifies its behavior on controlled cases. They document the reference population, perturbation assumptions and whether the method describes association or a causal intervention. User testing can establish whether the explanation improves a real judgment without encouraging overconfidence. Artifacts may include model behavior summaries, local explanation views and failure examples. The practitioner also checks that explanations do not expose confidential features or suggest actionable changes that the model would not actually respond to.

Interpretability does not establish fairness, correctness or causal validity. Local surrogates can be unstable and feature attributions depend on a baseline or treatment of correlated inputs. An explanation can be accurate yet unusable for its audience, or persuasive yet inaccurate. Quality checks should ask what decision the explanation improves and whether that improvement survives realistic cases, including errors and situations outside the model's supported domain.