Commercial AI-based Mental Health Chatbots as Low-Intensity Adjuncts to Psychotherapy: Effectiveness, Adherence, and Safety—A Systematic Review and Meta-analysis

INTRODUCTION: The effectiveness, scalability, and safety monitoring of commercial direct-to-consumer AI mental-health chatbots remain uncertain. METHODS: In this systematic review and meta-analysis, eight databases were searched for English peer-reviewed studies (Jan 1, 2015-Dec 31, 2025). RCTs were meta-analyzed with random-effects models (Hedges' g). Completion was synthesized with a one-stage binomial-logit generalized linear mixed model; attrition was pooled as risk ratio. Safety focused on harms-reporting interpretability. RESULTS: Fifty-two studies (22 RCTs; N = 110,594) covering 13 commercial chatbots were included. Depression improved versus controls (18 RCTs; N = 3,170): g = -0.35 (95% CI -0.56, -0.13), about 1.6 PHQ-9 points, below the 5-point MCID; NNT about 6.6. Anxiety effects (15 RCTs; N = 2,936) were uncertain: g = -0.37 (95% CI -0.87, 0.12). Loneliness partially reduced (4 RCTs; N = 662): g = -0.21 (95% CI -0.39, -0.03). Completion pooled at 79.5% (95% CI 68.0, 87.6%) but showed a scalability gradient. Attrition was similar to controls (RR = 1.01; 95% CI 0.79, 1.29). Harms reporting was uninterpretable: only 3/22 RCTs met active-ascertainment criteria. Certainty ranged from low to very low. CONCLUSION: Commercial AI mental-health chatbots may yield modest improvements most consistently for depression. Loneliness evidence is limited and fragile, while anxiety and affect outcomes remain heterogeneous; for anxiety, net harm cannot be excluded. Completion rates likely overestimate real-world retention, and safety cannot be inferred from "no events" claims in the absence of harms monitoring. These products are best positioned as low-intensity adjuncts requiring safety-by-design baselines.