Adaptive “Surprise-Driven” Explanations: Learning from User Deviations

Most explainable AI systems assume users’ needs are static or can be captured by broad personalization, but as Chen et al. (2023) and Das et al. (2021) highlight, users often react in unexpected ways—sometimes distrusting explanations or overlooking key information. This idea proposes instrumenting XAI interfaces to monitor user behavior (e.g., hesitation, repeated queries, disengagement) as proxies for “surprise” or confusion. When such deviations are detected, the system would generate targeted follow-up explanations, offering alternative perspectives (e.g., visual, rule-based, counterfactual) or clarifying context-specific details. This builds on the “investigate deviations” heuristic and aligns with Ribes et al. (2021) and Panigutti et al. (2022) who note the importance of user perceptions but goes further by creating a feedback loop between real user behavior and explanation strategy. The result could be higher user trust and satisfaction, especially in high-stakes domains like healthcare or finance, where misunderstanding an explanation can have major consequences.

References:

1. Explainable AI for Robot Failures: Generating Explanations that Improve User Assistance in Fault Recovery. Devleena Das, Siddhartha Banerjee, S. Chernova (2021). IEEE/ACM International Conference on Human-Robot Interaction.
2. When Post Hoc Explanation Knocks: Consumer Responses to Explainable AI Recommendations. Changdong Chen, Allen Ding Tian, Ruochen Jiang (2023). Journal of Interactive Marketing.
3. Understanding the impact of explanations on advice-taking: a user study for AI-based clinical Decision Support Systems. Cecilia Panigutti, Andrea Beretta, F. Giannotti, Dino Pedreschi (2022). International Conference on Human Factors in Computing Systems.
4. Trust Indicators and Explainable AI: A Study on User Perceptions. Delphine Ribes, Nicolas Henchoz, Hélène Portier, Lara Défayes, Thanh-Trung Phan, D. Gática-Pérez, A. Sonderegger (2021). IFIP TC13 International Conference on Human-Computer Interaction.