Therapeutic Interaction Features of AI Chatbots in Depression Interventions: Systematic Review and Meta-Analysis

Depression is a prevalent mental health disorder and a leading cause of disability worldwide, creating substantial personal and societal burdens. Digital mental health interventions have emerged as accessible and scalable solutions, with artificial intelligence (AI)-driven chatbots increasingly applied to deliver therapeutic content, monitor symptoms, and provide personalized support. However, limited evidence exists on how chatbot interaction features influence treatment adherence and clinical outcomes in depression.

This systematic review aimed to evaluate the clinical effectiveness of AI-driven chatbots for depression and to examine the associations between chatbot characteristics, treatment outcomes, and user adherence.

A systematic review and meta-analysis were conducted following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines, searching 6 databases (Web of Science, Scopus, PubMed, IEEE Xplore, Embase, and APA PsycINFO) for randomized controlled trials (RCTs) published before May 30, 2025. Eligible studies involved individuals with depression or depressive symptoms receiving AI-driven chatbot, conversational agent, or virtual therapist interventions, with outcomes assessed using the Patient Health Questionnaire-9 (PHQ-9). Data extraction included chatbot type, interaction features, adherence, and standardized mean differences (SMDs) for symptom change. Risk of bias was assessed using the Cochrane Risk of Bias tool version 2 (RoB 2). Random-effects meta-analyses were performed with the Hartung-Knapp-Sidik-Jonkman adjustment. This review was preregistered on the Open Science Framework.

A total of 11 RCTs involving 2220 participants (1091 in the intervention and 1129 in the control groups) were included. Using a random-effects model with Hartung-Knapp-Sidik-Jonkman adjustment, AI-driven chatbots showed a small-to-moderate reduction in depressive symptoms compared with control conditions, but the effect was not statistically significant (SMD=-0.46, 95% CI -1.02 to 0.10; P=.01; 95% prediction interval -1.50 to 0.58). Subgroup analyses of adherence did not show significant differences across the reported chatbot-type subgroups. In contrast, exploratory analyses of interaction features revealed more consistent patterns for adherence. Emotional responsiveness, structured feedback strategies, and interaction frequency were associated with higher adherence in high-scoring subgroups, whereas dialogue depth, self-disclosure encouragement, and user agency level showed weaker or inconsistent associations. For clinical outcomes, associations with interaction features were less consistent and more heterogeneous.

This systematic review provides an interaction-focused synthesis of AI-driven chatbot interventions for depression, examining how interaction features relate to clinical outcomes and user adherence. Although overall effects were not statistically significant, emotional responsiveness, structured feedback, and interaction frequency were consistently associated with higher adherence. Engagement and outcomes may be influenced by distinct mechanisms. Limitations include the small number of RCTs, heterogeneity, reliance on study-reported descriptions, and potential publication bias. These findings highlight the importance of interaction design in developing scalable digital mental health interventions.