Explainability & Traceability

Methods like SHAP and LIME that attribute model predictions to input features. These enable deployers to explain which factors drove a specific decision — essential for Article 13 transparency obligations and GDPR's right to explanation.

Visualising attention weights in transformer-based models to show which parts of the input the model focused on. Useful for natural language and image models, but attention does not always equal explanation — context matters.

Explaining decisions by showing what would need to change for a different outcome. 'Your application was declined because X — if X were Y, the outcome would be different.' The most actionable form of explanation for affected individuals.

Structured documentation of model purpose, performance metrics, limitations, ethical considerations, and intended use. Model cards are the bridge between technical performance and regulatory compliance — required for both GPAI and high-risk systems.

Explanations in terms of human-understandable concepts rather than raw features. These map model behaviour to domain concepts that regulators and affected parties can meaningfully evaluate.

Communicating the model's confidence level alongside its prediction. High-risk decisions made with low model confidence should trigger human review — connecting explainability directly to human oversight requirements.