Features of Explainability: How users understand counterfactual and causal explanations for categorical and continuous features in XAI

Counterfactual explanations build trust but may not improve actual understanding: While counterfactual explanations (which show "if X had been different, the outcome would change") made users feel more satisfied and trusting compared to causal explanations, they did not help users actually predict AI decisions any better, suggesting a gap between perceived and real comprehension.

Causal explanations are as effective as counterfactuals for decision understanding: In a study of 127 participants, causal explanations performed equally well as counterfactual explanations at helping users accurately understand why an AI system made a particular decision, indicating that simpler, rule-based explanations may be underutilized.

Categorical features are easier to explain than continuous ones: Users understood explanations about categorical features (like "employment status: yes/no") significantly better than explanations about continuous features (like "income: $45,000"), suggesting that current explanation methods may need different approaches depending on feature type.

User satisfaction doesn't guarantee genuine understanding: The study reveals an important disconnect—people reported higher satisfaction with counterfactual explanations even though these explanations didn't improve their actual ability to understand AI decisions, raising concerns about whether users are truly informed when they feel confident.

Current counterfactual methods may need refinement: The research challenges the assumption that counterfactual explanation methods work equally well across all types of data features, suggesting developers should reconsider how these explanations are designed and presented to different audiences.