Background <p>The integration of AI-based conversational agents (CAs) in mental health aims to bridge the global “treatment gap.” However, high user attrition poses a critical ethical challenge, potentially compromising care continuity and distributive justice. This study quantifies attrition rates in AI-driven mental health interventions and evaluates their ethical implications for user autonomy and research integrity.</p> Methods <p>Following PRISMA 2020 guidelines and registered with PROSPERO, this systematic review and meta-analysis synthesized evidence from six databases through September 2025. Employing a convergent parallel mixed-methods design, we analyzed 39 independent studies. A random-effects model was used to calculate pooled attrition rates, while meta-regression explored the correlation between attrition and clinical efficacy (improvement in depressive/anxiety symptoms). Qualitative narrative synthesis was further applied to contextualize user experiences.</p> Results <p>The meta-analysis revealed a pooled attrition rate of 17.04% (95% CI, 11.46%-23.38%) for AI interventions. Generative AI-based CAs exhibited significantly lower attrition rates (9.25%) compared to retrieval-based systems (23.12%, <i>p</i> = 0.044). Subgroup analysis identified intervention target, delivery platform, interaction mode, engagement reminders and safety measures as key moderators of engagement. Crucially, a significant inverse association between retention and treatment efficacy was observed in depression interventions, where lower attrition correlated with larger effect sizes (β = 0.97, <i>p</i> = 0.043). Furthermore, a pervasive “reporting gap” exists regarding the qualitative reasons for user withdrawal.</p> Conclusions <p>High attrition rates in AI-based mental health interventions may represent more than technical disengagement, potentially implicating the ethical principles of beneficence and non-maleficence by disrupting the continuity of care. While generative AI shows promise in fostering a perceived therapeutic alliance and mitigating attrition, the observed association between disengagement and diminished efficacy suggests a risk of exacerbating structural health disparities. To address these challenges, we propose the “Ethics-oriented Engagement Framework”(EEF), which integrates transparent accountability, autonomy balance, safety-by-design, and inclusive beneficence. This framework provides a normative roadmap to reconcile the tension between promoting user engagement and upholding the foundational ethical mandates of patient safety and justice.</p> Trial registration <p>PROSPERO (CRD420251275051).</p>