Predicting Satisfaction of Counterfactual Explanations from Human Ratings of Explanatory Qualities

Feasibility and trust are the most critical factors in making AI explanations satisfying to users. The research found that when an AI system explains its decision by suggesting changes, people care most about whether those changes are actually doable and whether they would genuinely produce the desired outcome—more so than other qualities like how simple or complete the explanation is.

Current technical metrics for evaluating explanations are incomplete. Traditional measures like "sparsity" (how few changes are suggested) or "proximity" (how similar the changed scenario is to the original) don't capture what actually matters to humans. This gap means AI developers may be optimizing for the wrong things when designing explanation systems.

Multiple explanation qualities matter beyond the top two. Even when feasibility and trust are set aside, other factors like completeness, understandability, and coherence still explain 58% of why people find explanations satisfying. This suggests there's no single "magic ingredient"—responsible AI systems need to balance several different qualities.

More detailed explanations aren't automatically better. The study found that complexity (explanation length and detail) operates independently from satisfaction, meaning adding more information doesn't necessarily make explanations more helpful. This challenges the assumption that comprehensive explanations are always preferable.

Human evaluation data can improve AI explanation systems at scale. By analyzing ratings from 206 participants across seven explanatory criteria, the researchers created a predictive model that could assess explanation quality without conducting expensive user studies each time—offering a more practical path for responsible AI development.