Conversational AI in Therapy: Current Applications and Future Directions in Mental Health Support

This paper delivers a rigorous mixed-methods synthesis of conversational AI applications in mental health therapy, analyzing 47 randomized controlled trials, 19 quasi-experimental studies, and 11 real-world datasets totaling over 142,000 participants across 22 countries. Quantitative meta-analysis reveals moderate effect sizes (SMD 0.30– 0.45) for AI-driven interventions, comparable to low-intensity clinician treatments, particularly in CBT-based approaches for mild-to-moderate depression. Advanced NLP models achieve over 90% intent recognition accuracy, and personalized AI leveraging Big Five traits enhances user engagement and disclosure. However, significant challenges persist up to 25% performance drop in non-English contexts, inconsistent crisis protocols in only 40% of clinical settings, and pervasive data privacy concerns with 70% of users worried and only 30% of apps providing clear policies. Attrition rates average 14.7%, with notable dropouts among youngest and oldest users. Qualitative analysis of 4,200+ pages of transcripts and 10,700 feedback logs uncovers barriers including cultural mismatch, algorithmic bias, and limited long-term data beyond six months. Hybrid clinician-AI models show promise in improving adherence and reducing dropout. The study underscores the urgent need for large-scale, multicultural trials, comprehensive ethical frameworks, and culturally adapted AI to optimize safety, efficacy, and equity in digital mental health care globally. This work sets a benchmark for evidence-based deployment of conversational agents, highlighting both transformative potential and critical limitations in current AI-mediated mental health interventions.