Toward Affective XAI: Facial Affect Analysis for Understanding Explainable Human-AI Interactions

Emotional responses matter in AI explanations: Current explainable AI (XAI) research focuses on what systems explain but largely ignores how people feel when receiving those explanations. This paper demonstrates that facial expressions—specifically increased inner eyebrow raising (AU1) and brow lowering (AU4)—indicate when users are struggling to understand or effectively use AI explanations.

Facial expressions reveal comprehension problems: The research found that heightened arousal and specific facial muscle movements correlate with participants failing to use explanations effectively, suggesting that emotions can serve as real-time signals that an explanation isn't working for a particular user or task.

One-size-fits-all explanations are insufficient: The findings support personalizing explanations based on individual interaction styles and task difficulty. Rather than providing static explanations to all users, AI systems could potentially adapt their communication approach when they detect signs of user confusion or frustration through facial analysis.

Responsible AI needs human-centered design: This work highlights that truly responsible AI requires understanding the complete human experience—not just whether explanations are technically accurate, but whether they actually help people feel confident and capable when making decisions with AI assistance.

Facial affect analysis could improve AI accountability: By monitoring emotional responses during human-AI interactions, developers could identify where explanations fail and iteratively improve them, creating a feedback loop that makes AI systems more transparent and trustworthy in practice.