Designing LLM-Based Conversational Agents for Prosocial Decisions: The Roles of Warmth and Competence

As Conversational agents increasingly support morally consequential decisions, it remains unclear whether anthropomorphic design promotes prosocial behavior. Integrating self-humanization theory with the stereotype content model, we examine how anthropomorphic cues shape perceived warmth, perceived competence and, in turn, algorithm aversion and prosocial behavior. In a preregistered, incentivized between-subjects experiment, 521 participants interacted with either a human-like conversational agent, a computer-like conversational agent, or a static text-based control before allocating money to an animal-related charity. Anthropomorphic cues increased both perceived warmth and perceived competence. However, only perceived competence was significantly associated with lower algorithm aversion and higher donation amounts, though the two paths did not differ significantly. Exploratory mediation analyses indicated an indirect association of anthropomorphic design with donations through competence. These findings qualify the presumed primacy of warmth in prosocial contexts and identify perceived competence as a pathway through which anthropomorphic conversational agents may support prosocial behavior.