Simulated Interviews for Child Abuse Detection

The mean length of experimental sessions was 52.1 min (SD = 14.4). After completing the informed consent and demographic information forms, the participants read short instructions (see Supplementary material 2) on best practice in interviewing children and answered two questions to confirm their understanding of these instructions. Participants were asked to read the instructions again if they gave an incorrect answer to one or both of the two questions.
Participants conducted simulated interviews with randomly selected two avatars out of the 14 available. Before each of the interviews, the participants first read a background scenario (see Supplementary material 3 for an example) of the alleged case and answered two questions about their preliminary impression of the case before the interview: (1) the presence of abuse (“present” or “absent”) and (2) confidence in their assessment on a 6 point scale (“50%: guessing” to “100%: completely sure”). Participants were instructed to focus their questions on eliciting information to determine the presence or absence of sexual abuse. Otherwise, they were free to ask any questions without restrictions. Each interview lasted 10 min.
After the interview, participants were asked three questions about their conclusions based on the information obtained in the interview: (1) the presence of the abuse (“present” or “absent”), (2) confidence in their assessment on a 6 point scale (“50%: guessing” to “100%: completely sure”), and (3) a description of what, according to them, had happened to the avatar. Unfortunately, the answers to these questions were missing for 27 (64%) participants due to a system error. Therefore, the correctness of conclusions was not included in the statistical analyses.
Between the two interviews, participants received either no intervention or either feedback or modeling as a training intervention. As the feedback intervention, the participants were provided two types of feedback after the first interview: (1) feedback consisting of the outcome of the case and (2) feedback on the questions (two recommended questions and two not recommended questions) asked in the interview. For feedback concerning questions, the AI avatar chose questions randomly from the questions recorded during the interview and then provided automated feedback regarding them. The modeling intervention included (1) reading a series of learning points of good and bad questioning methods and (2) watching a total of four 2.5-min videos of good and bad interviews with both an abused and a non-abused avatar. The contents of the modeling intervention were the same as those in Haginoya et al. (2021) (link). Participants read the background scenarios leading to the alleged cases before watching the modeling videos of each avatar and read the outcomes of the cases after watching the modeling videos.