Optimization of Late Imagery Rescripting Research Using Generative Artificial Intelligence

Overview

The aim of the study is to examine the effect of imagery rescripting (ImRs) in the context of utilizing large language models (LLMs). Intervention will involve the prior presentation of the most aversive fragment of the memory, the so-called 'hotspot.' This intervention will allow for the replication of the effect described by Dibbets and Arntz (2016), according to which the prior activation of the most emotional element of a memory enhances the effectiveness of ImRs. The study is also significant due to another ongoing study in which a substantial number of participants have already been examined; however, due to the exhaustion of funds, it was not possible to utilize the remainder of the recruited sample. Investigating an additional condition will allow for a more complete utilization of the available participant pool and significantly increase the project's scientific value by comparing the traditional ImRs mechanism with its AI-generated version.

This pilot randomized controlled trial will explore the application of large language models (LLMs) in the development of personalized therapeutic interventions. The study will focus on the emotional and psychophysiological effects of listening to autobiographical scenarios based on participants' childhood experiences of parental criticism. All participants will be asked to recall and describe two critical and two neutral childhood memories. Based on this input, personalized scripts will be generated using Gemini, a large language model. Each script will be reviewed and, if necessary, modified by trained experimenters to ensure therapeutic coherence and alignment with imagery rescripting (ImRs) principles. On Day 1, all participants will listen to critical personalized scenarios during the laboratory session. The experimental group will listen to modified versions of the critical memory scripts, in which a therapist figure will intervene to address the child's unmet needs-an application of imagery rescripting. To assess physiological arousal, skin conductance will be continuously recorded throughout the session. After each scenario, participants will rate their emotional intensity and specific feelings (e.g., fear, sadness) on Likert scales. The group will receive the ImRs intervention after the initial scenario phase. One week later, all participants will complete follow-up questionnaires assessing generalized anxiety (GAD-7) and the frequency of intrusive thoughts related to the memories. In addition, a panel of licensed cognitive-behavioral therapists will evaluate the generated scenarios for therapeutic quality. Their feedback will be used to assess the acceptability and coherence of AI-assisted therapeutic scripts. The study will test the feasibility of using LLM-generated content in clinical settings and aims to determine whether such interventions can reduce distress and intrusiveness while eliciting measurable emotional and physiological responses Hypotheses: The criticism scenarios generated by the model will elicit fearful responses in all participants. The level of fearful reaction evoked by the AI-generated criticism scenarios will correlate with participants' baseline fear of failure. Participants in the ImRs group will report fewer intrusive thoughts and lower generalized anxiety levels one week after the intervention. Magnitude of Prediction Error (operationalized as a difference in SCL response between hotspot and intervention parts of the scenario) will correlate positively with a decrease in the number of intrusive thoughts. The magnitude of imagery difficulty during the rescripting part will correlate negatively with the magnitude of Prediction Error. Subjective efficacy of intervention will be predicted by working alliance.