In a pioneering study, researchers show how AI-generated questions not only feel natural to children but also outperform untrained human interviewers in accuracy—paving the way for smarter, more ethical child forensic interviews.
Research: Comparing the performance of a large language model and naive human interviewers in interviewing children about a witnessed mock-event. Image Credit: mayam_studio / Shutterstock
In a first-of-its-kind study published in the journal PLOS ONE, an international team of researchers led by scholars from New York University Shanghai and Åbo Akademi University in Turku, Finland, explored the potential of artificial intelligence to assist in sensitive child investigative interviews. The study compared how effectively a Large Language Model (LLM), specifically ChatGPT, and untrained human interviewers were able to interview children about a mock event they witnessed.
The findings suggest that AI-powered interviews may complement human efforts by improving the questioning style and efficiency of child interviews. The study involved 78 children aged 6-8 who watched videos depicting events with potential for misinterpretation. Researchers then conducted interviews using either questions generated by ChatGPT or by human interviewers. The interviews were designed to evaluate the ability of both groups to elicit accurate and detailed accounts.
The AI-powered questions adhered closely to recommended best practices, emphasizing open-ended prompts like "Tell me what happened." Overall, LLMs posed the same proportion of recommended questions as humans and elicited as much correct but less incorrect information compared to the naive human interviewers.
The study found that while human interviewers asked more questions overall, ChatGPT-generated queries followed professional guidelines more consistently. The AI-driven interviews also elicited more correct information per question and produced less false information than human-led interviews. Most children believed they were being interviewed by a human, indicating that AI-generated questions felt natural and engaging.
Lead researcher Pekka Santtila stated, "Our results indicate that AI could play a crucial role in enhancing the quality of child interviews by supporting human interviewers with real-time question suggestions. This could be especially valuable in contexts where specialized interviewer training is limited."
However, the study's authors caution that while the findings are promising, further research is needed. Future investigations will explore how AI might assist in real-time interviews involving complex, emotionally charged cases.
With continuous advancements in AI and machine learning, the potential applications for LLMs extend far beyond current capabilities. The study advocates for development and cross-disciplinary collaboration to refine AI's integration into forensic and legal investigations.
The research was conducted by a multidisciplinary team from New York University Shanghai, East China Normal University, Åbo Akademi University, and several other institutions globally, with funding support from New York University Shanghai's Start-Up Fund and Sundellska Stiftelsen.
Source:
Journal reference:
- Sun, Y., Pang, H., Järvilehto, L., Zhang, O., Shapiro, D., Korkman, J., Haginoya, S., & Santtila, P. (2025). Comparing the performance of a large language model and naive human interviewers in interviewing children about a witnessed mock-event. PLOS ONE, 20(2), e0316317. DOI: 10.1371/journal.pone.0316317, https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0316317