Mapping the AI Trust Landscape: Perception-Grounded Typology with Varying Predictors of Trust and Willingness to Interact

As artificial intelligence (AI) products diversify, it remains unclear whether trust mechanisms generalize across systems or vary by users’ construal of AI. In a video-based vignette study (N = 242), participants evaluated 11 AI systems ranging from disembodied assistants to highly anthropomorphic agents. Exploratory factor analysis and hierarchical cluster analysis identified four perception-based categories: Functional Machines, Social Companions, Anthropomorphic Agents, and Disembodied Assistants. Multilevel models showed that perceived competence and warmth predicted trust across categories. At the same time, anthropomorphism showed mixed effects: familiarity and animacy increased trust, whereas mind attribution and uncanniness decreased trust, with stronger negative effects for embodied agents. We further observed a dissociation between trust and willingness to interact, with willingness to interact more sensitive to visceral discomfort and category ambiguity. These findings indicate that trust in AI is category-dependent and that effective design should align social cues with users’ mental models.