Effectiveness of Conversational Agents for Depression and Anxiety: Systematic Review and Network Dose-Response Meta-Analysis of Agent Types, Dose, and Human Support

AI Summary

AI-enhanced agents with intensive human support showed greater symptom improvement than pure human digital guidance and multimodal chatbots (Hedges g ~ -4.3 to -4.5).
Goal-oriented and highly structured modular chatbots showed nonlinear dose-response, with possible enhanced effects at low-to-moderate doses; findings are preliminary.
Overall evidence certainty was very low due to risk of bias, indirectness, imprecision, and publication bias for depression, limiting confidence and recommendations.

TL;DR (AI Generated)

OBJECTIVE: This review systematically evaluated the comparative effectiveness of various CA modalities and levels of human support, as well as potential dose-response relationships across different CA types.

Where data permitted, model-based network dose-response meta-analysis (MBNMA) was performed to examine dose-response associations.

NMA showed that AI-enhanced intensive human support (AEIHS) appeared to produce greater symptom improvement for depression and anxiety compared with pure human digital guidance (PHDG; Hedges g=-4.28, 95% CI -5.56 to -2.99) and open-ended multimodal-enhanced chatbot (OEMM; Hedges g=-4.49, 95% CI -6.31 to -2.66), whereas most other comparisons were not statistically significant.

Basic summary

J Med Internet Res. 2026 Nov 13;28:e104186. doi: 10.2196/104186.

ABSTRACT

BACKGROUND: Conversational agents (CAs) are increasingly used to deliver scalable interventions for depression and anxiety, but existing evidence has often treated these tools as a single category. The comparative effectiveness of different agent types, intervention doses, and levels of human support remains unclear.

METHODS: A comprehensive search of PubMed, Embase, Web of Science, Scopus, Ovid, CINAHL, and Psychology and Behavioral Sciences Collection (PBSC) databases was conducted from inception to December 2025. Eligible studies were randomized controlled trials (RCTs) evaluating CA-based interventions. We conducted a random-effects network meta-analysis (NMA) using postintervention outcomes and assessed inconsistency and risk of bias (RoB). Evidence certainty was assessed using the CINeMA (confidence in network meta-analysis) framework. Where data permitted, model-based network dose-response meta-analysis (MBNMA) was performed to examine dose-response associations.

RESULTS: A total of 41 studies (n=7577) were included. NMA showed that AI-enhanced intensive human support (AEIHS) appeared to produce greater symptom improvement for depression and anxiety compared with pure human digital guidance (PHDG; Hedges g=-4.28, 95% CI -5.56 to -2.99) and open-ended multimodal-enhanced chatbot (OEMM; Hedges g=-4.49, 95% CI -6.31 to -2.66), whereas most other comparisons were not statistically significant. MBNMA indicated that goal-oriented conversational agent (GOCA) and highly structured modular chatbot (HSMC) appeared to exhibit nonlinear effects, with preliminary evidence suggesting potentially enhanced response at low-to-moderate doses, though these findings require further investigation. However, publication bias was detected for depression (Egger P=.02), and the overall certainty of evidence was very low, primarily due to RoB, indirectness, and imprecision.

CONCLUSIONS: While hybrid CA models with enhanced human involvement appeared to show potential advantages over fully automated systems, the evidence base remains limited. The very low certainty of evidence constrains confidence in comparative rankings and dose recommendations. Future high-quality trials are needed to establish optimal CA architectures and intervention doses for scalable mental health care.

TRIAL REGISTRATION: PROSPERO CRD420251145207; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251145207.

PMID:43024932 | DOI:10.2196/104186

Document this CPD