A Typology of AI Anthropomorphism and Associated Risks in Human–AI Interaction: Evidence from Real-World Incidents

Human-like cues in AI can trigger anthropomorphism, facilitating human–AI interaction while also introducing risks. Yet existing research often examines these risks in fragmented ways and lacks a unified analytical framework grounded in real-world contexts. Drawing on cases from the Artificial Intelligence Incident Database (AIID), this study uses an iterative typology-construction approach to identify five forms of AI anthropomorphism: malicious, misplaced, misuse, excessive, and failed anthropomorphism. Structured along two analytical dimensions, locus of agency and intentionality, the typology explains how anthropomorphic cues participate in distinct harm pathways and reshape responsibility attribution across contexts. The findings highlight the context-sensitive and paradoxical nature of anthropomorphism: while it may facilitate interaction, it can also miscalibrate expectations, amplify vulnerability, and obscure accountability. Grounded in documented incidents, this study provides a structured analytical lens for the responsible design, use, and regulation of anthropomorphic systems by linking each category to specific governance leverage points.