Meta-analysis: AI chatbots are good for mental health, especially when no human butts in. Mental health disorders impose a substantial global burden, while access to care remains constrained. AI-based chatbots have emerged as scalable digital interventions, yet existing evidence is limited by heterogeneous intervention types and methodological shortcomings. To address these gaps, the present study conducts a systematic review and three-level meta-analysis to evaluate the effects of AI–based chatbots on mental health. AI architectures encompassed LLM-based systems, hybrid rule-based and machine learning approaches, and structured decision-tree systems Overall, AI-based chatbot interventions were associated with modest improvements across multiple mental health domains. No significant moderators were identified, although significant improvements were observed across depression, anxiety, loneliness, affect, and sleep. Neither interaction modality nor the integration of human support significantly moderated intervention effects. Notably, a counterintuitive pattern emerged: fully automated chatbot interventions showed significant effects, whereas those incorporating human support did not. Consistent with this, evidence from chatbot trials indicates that fully automated systems can produce meaningful improvements without therapist involvement. Taken together, these findings suggest that human support is not inherently beneficial in chatbot interventions, and its impact likely depends on its design, timing, and functional role rather than a simple guided versus unguided distinction. Several mechanisms may account for the positive but modest overall effect observed in the present meta-analysis. First, AI-based chatbots may reduce traditional barriers to mental health care, including high costs, long waiting times, limited provider availability, and stigma associated with help-seeking. By providing immediately accessible, anonymous, and geographically unrestricted support, these tools may reach individuals who might otherwise receive little or no intervention, thereby facilitating symptom improvement even in the absence of intensive therapeutic contact. Second, the conversational nature of chatbot interactions may enhance user engagement and adherence relative to static digital interventions. Unlike passive psychoeducational platforms, conversational agents deliver content through ongoing and responsive dialogue, which has been associated with greater user involvement and sustained engagement. Third, and perhaps most importantly, the observed benefits are unlikely to reflect the novelty of AI per se, but rather the delivery of established evidence-based therapeutic content through a more accessible medium.