Werewolf study: Pros' emotional control points to empathetic AI design

A new arXiv paper (arXiv:2608.04605) from Sho Mitarai, Chang Liu, Goshiro Yamamoto, and Nagisa Munekata examines how humans manage emotions in social deception games, using Werewolf as the testbed. The researchers tracked electrodermal activity (EDA)—a proxy for emotional arousal—during live gameplay, then analyzed communication patterns during spikes of physiological stress. They compared professional players (experienced in high-stakes deception) with nonprofessionals to uncover how expertise changes emotional expression and conversational strategy. The results showed a clear divergence: under high arousal, professional players leaned into persuasion-based language, maintaining their influence attempts, while nonprofessional players shifted toward information-focused communication, effectively changing their approach when stressed. Statistical analysis confirmed significant differences in expression patterns between the two expertise levels.

Beyond the gameplay dynamics, the study found that professional players exhibited far more stable emotional states throughout the match, indicating stronger emotional regulation under pressure. This matters beyond Werewolf—the findings directly inform how AI systems could model and respond to human emotional states. By mapping physiological signals to communication strategy adjustments, developers can build emotionally intelligent agents that detect when a user is stressed and adapt their tone, phrasing, or persuasive tactics accordingly. The paper, published in Entertainment Computing, suggests a blueprint for AI that doesn't just understand words but reads emotional context, a step toward more natural human-AI interaction in negotiation, customer service, and virtual assistants.