Artificial intelligence in criminal investigation: effects of large language model consultation on confirmation bias in a simulated investigation game

Objectives Confirmation bias can distort information search and evidence interpretation in criminal investigations. We tested whether consulting a large language model (LLM) could reduce this bias in group decisions.Methods In total, 108 participants formed 36 three-person teams randomly assigned to LLM consultation, theory-based information and debiasing strategies, or unaided control. Teams investigated homicide cases adapted from real cases in a web-based game. Outcomes were investigative cues opened, deliberation time, and culpability assigned to true perpetrators versus innocent suspects.Results Group assignment had no main effect on overall culpability accuracy or deliberation time. In weighted analyses, group interacted with case type: controls were less accurate in wrongful – than correct-conviction cases, whereas this difference was numerically smaller in the LLM and theory groups. The interaction was nonsignificant in unweighted analyses and should be interpreted cautiously. The LLM group opened more cues than controls in wrongful-conviction cases.Conclusions LLM consultation increased information search in wrongful-conviction cases but did not clearly improve accuracy. Theory-based support showed a numerically smaller weighted accuracy penalty, but no clear advantage of LLM consultation over theory-based support was established. These findings inform the design of AI-assisted investigative decision support.