New study reveals how children anthropomorphize LLM chatbots across 35 studies

A new systematic literature review from ACM IDC '26 analyzed 35 empirical studies published between 2022 and 2025 to understand how and why children anthropomorphize large language model (LLM) chatbots. Researchers Hansinie Madushika Jayathilake and Renkai Ma identified four primary drivers: human-like persona construction (chatbots mimicking names, emotions, or backstories), adaptive scaffolding (adjusting responses to a child's level), supportive companionship (consistent, non-judgmental listening), and non-human embodied design (e.g., robot avatars or animated interfaces). These drivers consistently led children to attribute human characteristics—like feelings, intentions, or consciousness—to non-human AI systems.

The review also cataloged five distinct outcomes of this anthropomorphism. Children exhibited paradoxical social and moral responses (e.g., being polite to a chatbot but also willing to hurt its feelings). They developed dual consciousness, simultaneously knowing the chatbot isn't real yet treating it as if it were. Interactions formed varied social ties—from temporary playmates to trusted confidants. Kids tested social boundaries by being rude or affectionate, and they often blamed or apologized to chatbots when conversations broke down. The authors highlight both benefits (emotional support, learning engagement) and risks (overattachment, privacy, manipulation), urging designers to create LLM chatbots that balance realism with clear non-human cues to support children's developmental needs.