XAI Handbook: Towards a Unified Framework for Explainable AI

Terminology chaos undermines XAI progress: The paper identifies that the explainable AI field lacks agreed-upon definitions for core terms like "explanation" and "interpretation," making it difficult to compare different methods or verify compliance with fairness and bias requirements—a critical gap for both scientific advancement and regulatory compliance.

A unified framework enables meaningful comparison: By proposing concrete definitions and standardized steps for producing explanations, the framework allows existing XAI methods (like LIME, SHAP, and MDNet) to be evaluated on comparable terms, revealing their specific advantages and limitations rather than treating them as interchangeable alternatives.

Standardization supports regulatory demands: The framework's structured approach to defining and measuring explanations helps meet regulatory requirements for AI transparency and fairness, providing a common language that organizations and regulators can use to assess whether AI systems meet accountability standards.

Not all explanations are created equal: The paper demonstrates that different XAI methods have different scopes and capabilities—meaning a method that works well for one type of AI system or use case may not be appropriate for another, requiring practitioners to understand these distinctions rather than applying tools indiscriminately.

Clear definitions matter for real-world deployment: By establishing what "interpretability" and "explanation" actually mean in concrete terms, the framework helps organizations move beyond vague commitments to "explainable AI" and instead implement specific, measurable approaches to making their AI systems more transparent and trustworthy.