AI Explainability & Interpretability Vocabulary

Interpretable models (logistic regression, decision trees) are intrinsically understandable. Black-box models (deep neural networks) require explainability techniques (SHAP, LIME) to provide post-hoc explanations of individual predictions.

SHAP values are grounded in Shapley values from game theory. Example: 'The age feature contributed +0.3 to the credit score prediction, while the high debt-to-income ratio contributed -0.8 — the primary reason for rejection.'

LIME is local: it explains one prediction, not the whole model. It samples nearby inputs, observes the model's behaviour in that local neighbourhood, and fits a simple interpretable model to approximate the complex model's local decision boundary.

Saliency maps (and related Grad-CAM, gradient-weighted class activation maps) highlight what the model 'looked at' when making its decision. Example: a diagnostic AI highlighting the tumour region that drove its classification.

Effective AI explanations for non-technical audiences use natural language, concrete examples, and avoid statistical jargon. They should answer the specific question the stakeholder cares about ('why was this decision made?') rather than explaining the method.