Pre-interview hypothesis generation: large language models (LLMs) show promise for child abuse investigations

Investigative interviews in child abuse (CA) cases are vulnerable to biases that can compromise the investigation's objectivity. Children's initial statements are often ambiguous or incomplete due to developmental and motivational factors, leaving room for adult interpretations and suggestive influences. These biases need to be minimized, particularly before the investigative interview, which is often the primary evidence. Although there is a recognized need for formulating case-specific hypotheses to guide interviews, practical guidance is limited. We compared the capabilities of Large Language Models (LLMs) and humans with varying expertise levels in generating pre-interview hypotheses. Participants included CA investigation experts (n = 21), psychologists (n = 60), naive participants (n = 60), and two LLMs (GPT-4 Turbo and Llama 2). Each human group generated hypotheses for four of eight case vignettes, while LLMs generated hypotheses for all eight. This resulted in 81 sets from experts, 240 sets from other human groups, and 480 sets from LLMs. Hypotheses were evaluated for testability, specificity, and comprehensiveness. GPT-4 outperformed humans in generating a higher number of hypotheses that were also more specific and comprehensive. This highlights the potential of LLMs to assist in hypothesis formulation, suggesting they can enhance the objectivity and thoroughness of CA investigations.