To interpret or to explain: towards a generalframework for explainable artificial intelligence

Abstract

As artificial intelligence (AI) systems increasingly shape high-stakes decisions, the demand for transparent and trustworthy behaviourhas grown correspondingly. Yet despite extensive research on explainableAI (XAI), the foundational concepts of interpretability and explainabilityremain ambiguous and often used inconsistently across disciplines. Thisconceptual fragmentation limits our ability to formulate rigorous expla-nation objectives, compare explanation methods, and evaluate their suit-ability in practical systems. This paper addresses these issues throughthree contributions. First, we provide a unified conceptual clarificationof interpretation (sense-reading) and explanation (sense-giving), drawingon insights from linguistics, philosophy, cognitive science, and knowledgemanagement. These definitions disentangle the cognitive, algorithmic,and communicative aspects of explanations in AI. Second, we introducethe Interpret/Explain Schema (IES), which specifies how interpretationand explanation arise within the data–model–output pipeline of an AIsystem. Third, building on the IES, we propose the General Frameworkfor Generating Explanations (GFGE), a modular and model-agnosticframework that organises the components required to construct expla-nations, regardless of model class or explanation technique. We validateGFGE by instantiating it with a broad range of XAI methods, includ-ing post-hoc attribution techniques, attribution-driven hybrid methods,counterfactual explanations, surrogate models, prototype and concept-based approaches, and intrinsically interpretable argumentation-basedmodels. These instantiations demonstrate that GFGE captures the struc-tural backbone shared across heterogeneous XAI techniques, offering aunifying and theoretically grounded foundation for designing, analysing,and comparing explanations in AI systems.

Affiliated Institutions

• Université Paris 8 FR

Related Publications