Impact of Companion Agent's Self-Disclosure Levels on User Perception

Abstract

Self-disclosure, the deliberate sharing of personal information, is central to building trust and relational closeness, yet its role in companion virtual agents (CVAs) remains unclear. Prior work suggests that when agents disclose, users perceive them as warmer, more trustworthy, and more engaging. We conducted a three-day within-subjects study where 36 adults interacted with an embodied CVA across three disclosure conditions: none, moderate, and high. Standardized questionnaires assessing trust, appropriateness, affect, and engagement revealed no significant differences across conditions. However, conversation log analysis showed that users reciprocated agent disclosure at deeper intimacy levels, with moderate disclosure producing the strongest reciprocity patterns. Qualitative feedback revealed that nearly half of participants rated moderate disclosure most authentic, while high disclosure raised credibility concerns, exposing an authenticity paradox. These findings demonstrate that self-disclosure shapes user behavior through implicit relational dynamics that standardized scales did not detect, establishing moderate disclosure as an effective baseline for CVA design.