Effect of an AI agent trained on a large language model (LLM) as an intervention for depression and anxiety symptoms in young adults: a 28-day randomized controlled trial (Preprint)

This is the authors' abstract. We don't add key points for this paper.

BACKGROUND Internet- and mobile-based psychological interventions (IMIs) have the advantages of being more accessible, less costly in terms of time and money, and more user friendly than traditional psychological interventions. Young adults face emotional problems in their daily lives that cannot be ignored, and youth are becoming the mainstay of the mobile internet. Large language modeling techniques are rapidly evolving and show great potential for understanding and generating natural language, enabling more effective interaction with human dialog than traditional techniques, and large language model (LLM)-based AI conversational agents may play an important role in intervening in young adults’ negative emotions. OBJECTIVE The aim of this study is to investigate whether an AI agent trained on a LLM can intervene in depression and anxiety in young people with negative emotions and to determine the effectiveness of such intervention. METHODS This study is a 28-day randomized controlled trial (RCT). Residents were randomly assigned to an intervention group or a waiting group according to the order in which they were successfully contacted by the staff, and each user was asked to engage in a total of 28 days of dialog intervention with the Douyin companion bot and complete three psychological questionnaires (on Days 1, 14, and 28); however, the intervention group began to receive the dialog intervention after completing the first questionnaire, and the waiting group began to receive the dialog intervention after completing the third questionnaire. During the first four weeks, the waiting group was treated as a blank control. The two groups of subjects completed the three questionnaires at exactly the same point in time. Each user's depression, anxiety, and positive and negative emotions were measured using the Patient Health Questionnaire (PHQ-9), the Generalized Anxiety Disorder Scale (GAD-7) and the Positive and Negative Affect Schedule (PANAS), respectively. RESULTS Data from 657 users who completed pre-, mid- and posttreatment measures were included in the analyses. The prevalence of mild depression and anxiety among participants at baseline was 44.3% and 23.4%, respectively. The results showed a significant reduction in negative moods in the intervention group after four weeks of the dialog intervention, which was not found when the dialog intervention was conducted for a fortnight. Repeated-measures ANOVA showed that the dialog intervention significantly reduced depression in the intervention group at two weeks and significantly reduced both depression and anxiety in the intervention group at four weeks. CONCLUSIONS Overall, LLM-based AI conversational agents can effectively alleviate the mild anxiety and depressive symptoms of young adults with negative emotions through dialog interventions.