Explainable AI

Techniques that make AI decisions understandable to humans, showing how and why a model produced an outcome.

Also known as: XAI

In plain English

Explainable AI means an AI system shows why it made a decision, like a teacher explaining a grade. Think of it as a dashboard that reveals the key reasons so you can trust or challenge the result.

In practice

In practice, explainability appears as model cards, feature-attribution reports, counterfactual examples or audit logs when buying or deploying AI tools. It matters for regulatory compliance, incident investigation, user trust and faster debugging of models in production.

Under the hood

Explainable AI covers intrinsic interpretable models (for example decision trees or linear models) and post-hoc methods such as feature‑attribution (SHAP, LIME), surrogate models, saliency/attention visualisations and counterfactual explanations. Explanations are assessed by fidelity, stability and usefulness to stakeholders, and engineers must balance explanation simplicity against fidelity to the original model; attention weights are not by themselves reliable causal explanations.

Example

Please include explainable AI outputs and feature‑attribution reports so compliance can audit decisions and users can request reasons.

Related terms

AI Bias · Human-in-the-loop · Responsible AI · Evals

•