Scientific Explanation and Natural Language: A Unified Epistemological-Linguistic Perspective for Explainable AI

Explanations serve a unification function beyond simple logic: The paper challenges the common assumption that explanations work primarily through inductive or deductive reasoning. Instead, explanations unify disparate facts and concepts, meaning AI systems designed to explain should be built to connect and integrate information rather than just follow logical chains.

Causation and mechanisms are essential to meaningful explanations: For an AI explanation to be scientifically sound, it must identify the actual causes and mechanisms responsible for an outcome. This suggests that responsible AI systems cannot simply correlate inputs with outputs—they must trace the causal pathways that lead to decisions or predictions.

Natural language explanations require more than just technical accuracy: The research finds that effective explanations incorporate pragmatic elements like definitions, properties, and taxonomic relationships alongside causal-mechanistic information. This implies that AI explanations designed for human understanding should go beyond bare technical details to provide context and definitions that make explanations accessible.

Theory should inform XAI design and evaluation: The paper emphasizes that current methodologies for building and testing explanation-generating AI systems lack grounding in established philosophical theories of explanation. This suggests the responsible AI community should invest in bridging academic philosophy and AI engineering to create more rigorous standards for what counts as a "good" explanation.