Using large language model based AI suspects to train strategic use of evidence: preliminary evidence of transfer to mock suspect interviews

Objectives: The Strategic Use of Evidence (SUE) is a technique that aims to improve the ability to differentiate between liars and truth-tellers. However, while theoretical training provides guidance on interview techniques, it lacks opportunities for practical application. Hypotheses: We developed two Large Language Model driven AI Suspects with whom participants could simulate interviews and hypothesized that these simulations would enhance the transfer of training to later interactions with Human Mock Suspects. Method: The study included 156 Chinese laypersons (78 Interviewers and 78 Human Mock Suspects). The two AI suspects followed response rules representing simplified and prototypical examples of liars’ and truth-tellers’ behaviors under the SUE model. Interviewers were randomly allocated to one of three types of training: (a) Instruction & AI Exercise, (b) Instruction, and (c) Control. After the training, the participants interacted with either a lying or truthful Human Mock Suspect. Results: Receiving interventions made Interviewers use Evidence Framing Matrix (EFM: an important tactic within the SUE framework) more frequently, thereby eliciting more inconsistencies between the lying Human Mock Suspects’ statements and the evidence (i.e., evidence-statement inconsistencies) as well as more inconsistencies within their own statements (i.e., within- statement inconsistencies). Both Instruction and Instruction & AI Exercise groups used evidence-statement (in)consistencies more to make their judgments about whether Human Mock Suspects were lying or truthful compared to those in the Control group. Additionally, the Instruction & AI Exercise group was better at accurately judging whether the Human Mock Suspects were lying or truthful compared to the Control group. Conclusions: Overall, this study provided preliminary evidence that simulated SUE training with AI Suspects transferred to interactions with Human Mock Suspects in a controllable experimental setting but that the advantage over instruction-only was not particularly robust.