AI chatbot advice did not clearly help 36 mock detective teams judge guilt

Group assignment had no main effect on overall accuracy in assigning culpability to true perpetrators versus innocent suspects, and no effect on deliberation time. Teams with the AI consultation opened more investigative cues than controls in wrongful-conviction cases, but that extra searching did not clearly translate into better judgments.

In weighted analyses, controls were less accurate in wrongful-conviction cases than in correct-conviction cases, while the gap was numerically smaller in the LLM and theory groups. That interaction was nonsignificant in unweighted analyses, so the authors urge caution, and no clear advantage of the LLM over theory-based support was established.

The experiment targeted confirmation bias, which can distort how investigators search for information and interpret evidence. Participants worked in three-person teams on group decisions, since investigations are often collaborative.

Teams were randomly assigned to consult an LLM, to receive theory-based information and debiasing strategies, or to proceed without help. They investigated homicide cases adapted from real cases in a web-based game, with some cases ending in a correct conviction and others in a wrongful one.

Controls were less accurate in wrongful-conviction than correct-conviction cases in weighted analyses, and this difference was numerically smaller in the LLM and theory groups. The authors say this finding should be interpreted cautiously because it disappeared in unweighted analyses.

The authors conclude that LLM consultation increased information search in wrongful-conviction cases but did not clearly improve accuracy, and say the findings inform the design of AI-assisted investigative decision support.