Are AI Experts Better Than AI Friends? Social Roles, Failure Signals, and Mental Factors in Task-Oriented Interactions with AI Conversational Agents

Social role assignment is a promising design strategy for AI conversational agents (CAs), yet how these roles operate under failure remains unclear. Drawing on the concept of communicator reward valence (CRV) within expectancy violation theory (EVT), we conducted a two-part 3 × 2 between-subjects experiment (N1 = 145, N2 = 185) examining how social roles (expert vs. friend vs. default) and failure signals (present vs. absent) influence users’ continuance intention. Study 1 found that the expert role amplifies the negative impact of failure and performs significantly worse than the friend role, thereby challenging the expert role’s presumed superiority. Study 2 introduces mental factors as mediators organized by the stereotype content model (SCM). Mediation analysis shows that the expert role’s disadvantage stems primarily from deficits in warmth-related and cross-dimensional factors, with a competence-related factor (expertise) exhibiting a masking effect. These findings suggest that warmth is more critical for building human-AI relationships resilient to failure.