How conversational styles relate to trust in virtual humans: Evidence from a synthetic dialogue corpus based on Media Equation Theory

Abstract

This study investigates how conversational styles in virtual humans shape social perceptions, trust, and downstream attitudinal outcomes. Drawing on Media Equation Theory and the warmth–competence framework, we developed a controlled synthetic dialogue corpus of 9720 dialogues generated across two large language model families and three interactional contexts (news, customer service, and education). Four conversational style factors—humor, minor mistakes with self-correction, formality, and friendliness—were systematically manipulated to assess their effects on warmth, competence, and human-likeness. Structural equation modeling revealed that humor and friendliness significantly enhanced warmth, minor mistakes with self-correction increased human-likeness, and formality improved competence while diminishing warmth. In turn, warmth, competence, and human-likeness each predicted trust, which was associated with increased persuasion and relational orientation. Sequential mediation analyses indicated that style effects were transmitted indirectly via social perceptions and trust, while multi-group comparisons identified context-specific pathways: formality was most influential in news contexts, humor and friendliness in customer service, and self-correction in education. Together, these findings refine theoretical accounts of human–AI interaction by demonstrating that distinct stylistic cues selectively activate different dimensions of social cognition, and that trust formation depends on balanced, contextually congruent style configurations. Methodologically, the study demonstrates the utility of a synthetic dialogue corpus as a scalable and reproducible approach for theory-driven trust research. Practically, the results provide design-relevant insights for virtual news anchors, service chatbots, and educational tutors that must balance professionalism, warmth, and authenticity, while highlighting the need for complementary validation with human users. © 2026 The Author.