83. Exploration of the healing mechanism of virtual digital humans empowering patients with a history of psychiatric conditions

Abstract Background Currently, individuals with lived experience of mental illness face significant barriers in accessing mental health services, including persistent stigma, insufficient support resources, and limited accessibility of traditional service models. In recent years, digital mental health interventions have developed rapidly. Among them, Artificial Intelligence (AI) avatars, characterized by their high accessibility, interactive anthropomorphism, and emotional responsiveness, offer new possibilities for bridging existing service gaps. However, existing research has predominantly focused on efficacy verification, leaving a lack of in-depth exploration into the specific psychological and social mechanisms through which they exert therapeutic effects in this population. Therefore, this study aims to empirically and systematically analyze the potential therapeutic pathways and mechanisms of action of AI avatars in providing mental health support to this group. Methods The study recruited 96 participants who self-reported lived experience of mental illness and current moderate psychological distress. They were randomly assigned to Group A (n = 48) and Group B (n = 48). Group A engaged in structured dialogue support via an interactive interface with an AI avatar trained on principles of empathy and cognitive-behavioral therapy, for 20 minutes twice a week over a period of 6 weeks. Group B received standardized mental health education text materials of equal frequency and duration. Assessments were conducted before and at the end of the intervention using the Kessler Psychological Distress Scale (K10), the University of California, Los Angeles Loneliness Scale (UCLA-LS), and a self-developed Working Alliance Inventory (measuring the sense of connection with the avatar). Repeated-measures analysis of variance was used to test the intervention effects, and path analysis was employed to preliminarily explore the mediating role of the working alliance. Results Repeated-measures ANOVA showed a significant group (A vs. B) × time (pre- vs. post-intervention) interaction for K10 scores (F (1, 94) =21.37, p<.001, η2 = 0.19) and loneliness scores (F (1, 94) =18.52, p<.001, η2 = 0.16). Group A demonstrated a significant reduction in K10 scores (mean difference = −5.82, p<.001) and loneliness (mean difference = −4.15, p<.001) after the intervention, whereas no significant changes were observed in Group B. Path analysis indicated that the working alliance formed with the AI avatar significantly mediated the alleviation of psychological distress (indirect effect β = −0.31, p=.002). Discussion The study confirms that structured interaction with an AI avatar can effectively reduce psychological distress and loneliness in individuals with lived experience of mental illness, and that the therapeutic effect is partially achieved through the establishment of a positive working alliance. This suggests that the therapeutic mechanism of AI avatars may stem from providing a safe, accessible, and non-judgmental interactive space, operating through pathways of emotional resonance and cognitive restructuring. The findings provide mechanistic evidence for developing digital therapeutic tools targeting mental health vulnerable groups. Future research should focus on the long-term maintenance of effects, the impact of the avatar's level of anthropomorphism, and further exploration of its differential pathways of action across various subgroups. Funding No. KT2510097.