Testing large language model capability in building rapport and interviewing children about a witnessed mock event

This two-experiment study compared Large Language Models (LLMs) and humans in interviewing children about mock-events, examining rapport-building effectiveness and question presentation modes (human/avatar). In Experiment 1 (LLM/human-generated rapport questions with standardised memory queries), LLM interviews elicited greater verbal engagement and more detailed responses from children compared to humans. Though children’s interviewer liking remained equivalent, LLM-led rapport-building enhanced subsequent recall accuracy during memory phases. Experiment 2 reversed roles (standardised rapport, interviewer-generated memory questions), revealing LLMs asked fewer questions overall and used fewer recommended open-ended prompts than trained human interviewers. However, individual LLM-formulated questions extracted more correct details per question, albeit with increased central-detail errors. Question-presentation mode showed no significant effects in either experiment. Results demonstrate LLMs’ potential to boost engagement and memory recall through rapport-building strategies but highlight limitations in autonomous question formulation quality. While AI systems could enhance investigative interviews by supplementing human-led processes, current findings caution against fully autonomous implementations. Future research should optimise hybrid human-AI approaches to balance engagement benefits with error mitigation in child witness interviews.