DEMO
Koushiki Pohit, Magzhan Mukanova, Metali Mangal, David M Whittinghill
Purdue University
This pilot study examines how the three emotional states, calm, nervous, and aggressive, when expressed by conversational agents, persuade participants in a murder mystery game implemented in Unity Game engine. Using a within-subjects design, participants in our pilot study interacted with all three LLM agents, each designed as a suspect in a murder mystery narrative, with participants taking the role of a detective questioning them to determine the murderer. Each LLM agent was provided with a role, their relationship with the victim and alibis along with their respective emotional state. The Russell's valence-arousal scale showed that participants could distinguish between aggressive and calm agents, but the distinction between nervous and calm agents was less clear. Persuasiveness was measured using the 9-item Perceived Persuasiveness Scale (9PPS), which encompasses effectiveness, quality, and capability. Non-parametric analyses revealed a significant effect of emotional state on perceived effectiveness, with post hoc comparisons indicating that the calm agent was significantly more effective than the aggressive agent. These results suggest that emotional tone selectively influences persuasive effectiveness, and that calm emotional expression may be more effective than aggressive expression in conversational agents.