Developing an AI-driven Child Avatar Chatbot to Improve Interviewing Skills for Professionals Investigating Suspected Child Abuse and Neglect

Abstract

Children who come for investigative interviews due to being subjected to alleged child abuse or neglect need to meet professionals who conduct high-quality interviews. In child protection services (CPS) and law enforcement, high-quality interviews are the best means of detecting and investigating child abuse and neglect. Research has found that children aged 3-4 years can provide relevant information about experienced abuse in an interviewing context. Established research-based interview guidelines recommend the exhaustive use of ageappropriate, open-ended questions before turning to more specific cued-recall and directive questions, avoiding leading questions. Still, it has been found that open-ended questions are rarely used when interviewing children. The complex skills of interviewing require specialized training. Training programs of various intensities have been implemented worldwide. The use of research-based guidelines including a hypothesis-testing approach and interview training have focused on law enforcement contexts, however, awareness within the child welfare system has increased. Despite implementation of training, field studies find discrepancies between recommendations and practices, thus identifying a transfer-of-skills problem. Research has identified effective elements of interview training and has found that ongoing and realistic practice of skills are critical elements. Various simulation approaches for staging practice have been used in existing programs, such as mock interviews and simple avatars, but their use is resource demanding. Recent developments in advanced technology within the area of artificial intelligence (AI) have motivated research groups to develop avatar solutions. Target services, such as the CPS and police, have busy schedules and limited resources for training. The aim of this thesis was the research and development of an AI-driven child avatar prototype for realistic practicing of skills based on principles of human learning. Using quantitative approaches, the research had four aims. The first aim was to add knowledge on how frontline professionals interact with a digital child avatar, LiveSimulation (LiveSim), to train questioning skills from a user engagement perspective. The second aim was to assess and compare the questioning style in two frontline contexts, the CPS and the police, conducting child interviews following the same interviewing method. The third aim was to expand knowledge about preschool-aged children’s disclosure patterns and capacity to report in forensic interviews concerning suspected abuse. The fourth aim was to test whether training questioning skills by conducting multiple interview sessions using an AI-driven child avatar chatbot (text-only) enhanced trainee interviewing proficiency related to interview quality. The AI-driven language model, Generative Pre-trained Transformer 3 (GPT-3), was trained and fine-tuned using interview data, and a feedback function for classifying questions was integrated. Study I was an exploratory field study that revealed that the online child avatar training tool, earlier shown to facilitate learning of open-ended questioning, evoked good user engagement. Trainees were mostly professionals using the avatar as a stand-alone activity or as part of an online training program. Engagement was measured across various activities, such as the number and timing of activity completion. Study II and III were field studies analysing real-life investigative interview transcripts. Study II showed that the use of open-ended questions was sparse among trained interviewers, both within the CPS and police. Both groups were found to rely mostly on directive questions, followed by non-recommended option-posing questions (including yes/no questions). CPS workers asked more option-posing than the police, increasing the probability of erroneous responses, but fewer suggestive questions which may risk tainting the child’s testimony. Study III confirmed the infrequent use of open-ended questions in forensic interviews with preschoolaged children. In a selected sample where all the children reported abuse-related information during the interviews, a wide variation was revealed in the timing of the initial disclosure. Onethird disclosed in the pre-substantive phase. Furthermore, children aged 3 provided forensically relevant information across multiple question-response interactions, comparable with the 5- year-olds. However, the preschool-aged children were interviewed using techniques that were leading and involved lengthy sessions, which did not align with best practices. This may raise questions about the validity and representativeness of the findings. Study IV found that interviewing a child avatar chatbot trained to be a 6-year-old allegedly abused girl improved the interview quality regarding recommended questioning, especially when combined with direct feedback provided by an automated feedback function. The feedback function was validated and found to be reliable for classifying questions. This thesis provides insights into various aspects of developing a real-time conversational child avatar. The initial results provided support for the further development of online avatar training solutions, as professionals within demanding working contexts showed engagement through training. Evaluative research on interviews conducted within the CPS and the police highlights the need for practitioners interviewing children to practice and refine these skills to ensure effective skill transfer into real-world practice. As a potentially resource efficient practice tool with integrated direct feedback, the self-run text-based child avatar demonstrates promise in enhancing interviewing skills supporting continued development.