From scientific theory to duality of predictive artificial intelligence models

Summary In studies employing explainable artificial intelligence (XAI), model explanation, interpretation, and causality are often not clearly distinguished, leading to potential misunderstandings of model performance or relevance. For predictive AI models used in the natural sciences, the path leading from model explanation and interpretation to causal reasoning is of particular importance because it bridges theory and hypothesis-driven experimental design. Selected concepts from scientific theory can be taken into consideration to generate a conceptual framework for putting predictions into scientific perspective and recognizing potential caveats. For explainable models, it is argued that the scientific rationale underlying model derivation plays a decisive role in assessing and understanding predictions and exploring causal relationships, giving rise to the notion of model duality, as introduced herein.