Conversational AI for mental health: a systematic review of effectiveness and design
-
Published: September 7, 2026
-
Page: 1555-1571
Abstract
This systematic review synthesised evidence on the applications, effectiveness, engagement, acceptability, and responsible design of large language model (LLM)- and AI-based conversational agents for mental health and psychological well-being. Following the PRISMA 2020 framework, Scopus, PubMed, and ScienceDirect were searched for English-language publications from 2023 to 2026. The search identified 388 records; 16 duplicates were removed, 372 records were screened, 56 full texts were assessed, and 21 eligible publications were retained. Because the corpus combined randomized and quasi-experimental studies with single-arm, qualitative, developmental, protocol, conceptual, and review publications, findings were synthesised thematically rather than pooled in a meta-analysis. Four themes emerged: effectiveness and outcomes; design and development; assessment and evaluation; and user perception, acceptability, and ethics. Controlled studies reported promising short-term improvements in depression, anxiety, stress, and well-being, while feasibility and qualitative studies indicated generally favourable acceptability but also concerns about engagement decline, privacy, emotional dependence, bias, transparency, and crisis management. The evidence supports conversational AI primarily as a supervised adjunct to mental-health care rather than an autonomous clinical decision-maker. Major limitations were heterogeneous designs and outcomes, limited long-term follow-up, geographic concentration, and incomplete reporting of screening reliability and search dates in the original review record. Future research should use adequately powered, multi-site trials, standardised outcomes, explicit safety benchmarks, and transparent human-oversight procedures.
This work is licensed under a Creative Commons Attribution 4.0 International License.