Human Emotions in AI Explanations

For adapting to humans, explainable AI (XAI) quality can benefit from human expressions beyond verbal expressions. This is particularly true for emotions, as emotions affect how information is processed and subsequent decision-making. We analyze how different explanation strategies of XAI are perceived when the human interaction partner shows high arousal and/or valence. In a between-subjects experimental setting, we show that individuals with low arousal follow advice with no attempt to any explanation. On the contrary, individuals in a highly aroused state respond best to explanations with some justification (guided explanations). Concerning varying levels of valence, we find no similar pattern. Our results suggest that specific XAI strategies should be adapted not only to humans’ cognitive needs but also to humans’ information processing capacities and needs, which depend on emotional arousal.