Explainability of AI system outcomes
Description
Provision of meaningful explanations for AI decisions appropriate to the audience and system risk level.
Implementation Guidance
Explainability Requirements
Define explainability requirements per AI system based on risk: high-risk systems (lending, medical, legal) require individual-level explanations, medium-risk systems require aggregate-level explanations, and low-risk systems require general system documentation.
Explanation Methods
Implement appropriate explainability techniques: feature importance (SHAP values, permutation importance), counterfactual explanations (what would need to change for a different outcome), rule extraction (simplified decision rules), and natural language explanations for non-technical audiences.
Audience-Appropriate Communication
Tailor explanations to the audience: technical stakeholders receive model-level explanations, business users receive decision factor summaries, affected individuals receive plain-language explanations of factors influencing their outcome, and regulators receive methodology documentation.
Evidence Requirements
- Explainability requirements by AI system risk level
- Implemented explanation method documentation
- Sample explanations for different audiences
- User comprehension testing results