PRISM: A Pragmatic Framework for Evaluating Counterfactual Explanations in XAI

Leila Methnani, Virginia Dignum, Andreas Theodorou

Keywords: Explainable Artificial Intelligence, Gricean Maxims, Human-Centric Explainable Artificial Intelligence

TL;DR: Evaluating counterfactual explanations through a Gricean lens.

Abstract: This paper introduces PRISM, a framework for evaluating counterfactual explanations in eXplainable Artificial Intelligence (XAI) through the lens of conversational pragmatics. By mapping evaluation metrics to Grice’s cooperative principles—quantity, quality, relation, and manner—PRISM highlights how explanation violations can encourage users to infer deeper meanings. Demonstrated through a dashboard applied to counterfactuals from income prediction data, PRISM emphasizes the social and interactive aspects of AI explanations, fostering iterative understanding and advancing human-centric XAI. Currently, we are refining the approach through qualitative studies.

Submission Number: 15