Explaining Groups of Instances Counterfactually for XAI: A Use Case, Algorithm and User Study for Group-Counterfactuals

Group explanations may be more useful than individual ones: The research introduces "group counterfactuals" that explain patterns across multiple similar cases rather than single instances. A user study with 207 participants found this approach produced "modest but definite improvements" in people's understanding of AI systems, suggesting that humans prefer coherent explanations covering multiple events rather than isolated case-by-case explanations.

Counterfactual explanations show promise for legal compliance: The paper notes that counterfactual explanations are "proposed" as compliant with regulations like GDPR, though the research itself focuses on their explanatory effectiveness rather than legal validation. This highlights why counterfactuals are gaining traction in responsible AI—they can show what would need to change for a different AI decision.

User studies reveal modest but real improvements in AI understanding: The controlled study measured both objective measures (accuracy) and subjective psychological measures (confidence, satisfaction, and trust). Results showed that group counterfactuals improved understanding, though the improvements were described as "modest," indicating that while this explanation method helps, no single explanation technique is a complete solution.

Explainability methods need to match how humans naturally think: The research emphasizes that counterfactual explanations align with "the contrastive nature of human explanation"—meaning people naturally understand things by comparing what happened to what could have happened. Group counterfactuals extend this by matching human preference for broad patterns rather than isolated examples.