Influence of prior and task generated emotions on XAI explanation retention and understanding

TL;DR

This study investigates how emotions modulate the retention and understanding of XAI explanations by separating task-unrelated prior emotions from emotion-like responses elicited during explanations. Using a multimodal measurement setup (HRV, facial expressions, EmoNet) and a six-phase interaction with a DSS (Flobi) in a Holt–Laury risk task, the authors assess verbal recall and GUI-based judgments of feature influence. Key findings show that task-unrelated emotions do not reliably affect retention, but explanation-induced arousal can negatively impact understanding, with certain personal attributes (e.g., gender, health status, political orientation) predicting retention. The work highlights the importance of managing emotional load in XAI to optimize user comprehension and suggests design considerations to mitigate excessive arousal during explanations.

Abstract

The explanation of AI results and how they are received by users is an increasingly active research field. However, there is a surprising lack of knowledge about how social factors such as emotions affect the process of explanation by a decision support system (DSS). While previous research has shown effects of emotions on DSS supported decision-making, it remains unknown in how far emotions affect cognitive processing during an explanation. In this study, we, therefore, investigated the influence of prior emotions and task-related arousal on the retention and understanding of explained feature relevance. To investigate the influence of prior emotions, we induced happiness and fear prior to the decision support interaction. Before emotion induction, user characteristics to assess their risk type were collected via a questionnaire. To identify emotional reactions to the explanations of the relevance of different features, we observed heart rate variability (HRV), facial expressions, and self-reported emotions of the explainee while observing and listening to the explanation and assessed their retention of the features as well as their influence on the outcome of the decision task. Results indicate that (1) task-unrelated prior emotions do not affected the ratantion but may affect the understanding of the relevance of certain features in the sense of an emotion-induced confirmation bias, (2) certain features related to personal attitudes yielded arousal in individual participants, (3) this arousal affected the understanding of these variables.

Figures (11)

• Figure 1: Study Setup during the first Hold and Laury decission (staged)
• Figure 2: Experimental procedure with six phases.
• Figure 3: Visual representation while Flobi was explaining which variables contributed to which risk type classification of the explainee. The red arrows indicated that the explainee's value of this variable contributed to an estimation of a lower risk type whereas a blue arrow indicated evidence for a higher risk type.
• Figure 4: Visualization of the user interface to answer the question what effect the user's value of the presented variable had on the system's estimation of the user's risk type. This task was used to measure the explainee's retention of each variable.
• Figure 5: SAM scores for Valence (left) and Arousal (right), indicating higher (= more positive) valence for participants in the Happy condition, and a trend to higher self-reported arousal for participants in the Fear condition.