← AI Terminology

Explainable AI - XAI

Explainable AI (XAI) is the field of methods and practices that make AI model predictions understandable to humans — explaining why a model made a specific decision, which features drove it, and how confident it is.

It is increasingly required by regulation (EU AI Act, GDPR's right to explanation) and demanded by users in high-stakes decisions.
Why It Matters in AI
A black-box model that says "loan denied" gives the applicant no recourse and the operator no way to audit for bias. XAI methods — SHAP, LIME, attention visualisation, saliency maps — open the black box enough to verify, debug, and trust AI decisions. Regulators are mandating explainability for high-risk AI systems; XAI tools have become a standard component of ML deployment pipelines.
Key Points
Aspect Description
LIME Approximates a black-box model locally with an interpretable model — works for any model
SHAP Assigns each feature a contribution value (Shapley value) — model-agnostic, theoretically grounded
Intrinsic Inherently interpretable models: linear regression, decision trees, rule-based systems
Regulatory GDPR Article 22: right to meaningful explanation for automated decisions
Saliency maps Gradient-based: pixels/tokens most responsible for the prediction
Attention maps Visualise where the model "looks" in an image or text — limited as an explanation (debated)
Simple Analogy
A loan officer who can say "I denied this loan because debt-to-income ratio was 45% (over our 40% threshold) and employment history was only 6 months" is explainable. A neural network that says "denied" with no reason is not — XAI retrofits that explanation capability onto models that don't have it natively.
Common Usage Examples
  • shap.Explainer(model)(X_test) — SHAP values for any model; shap.waterfall_plot() for individual predictions
  • lime.lime_tabular.LimeTabularExplainer — local surrogate model for a single prediction
  • Grad-CAM: saliency map showing which image regions activated the classification decision
  • Integrated Gradients: captum.attr.IntegratedGradients(model).attribute(input, target=class_idx)
  • SHAP summary plots: shap.summary_plot(shap_values, X) — global feature importance across all predictions
Summary
In short: XAI makes AI decisions interpretable — essential for debugging bias, building user trust, satisfying regulators, and making high-stakes automated decisions auditable.