Research paper · RP37

No Symptoms to Reduce

Background. Prior pooled estimates show that conversational agents reduce depressive symptoms in mixed samples (g = 0.31) but not in non-clinical adults (g = 0.07). Whether the same agents raise wellbeing, skill use or behaviour in people who are not ill has not been settled. Methods. We conducted a PRISMA-2020 systematic review and random-effects meta-analysis of randomised and controlled trials published from 2017 to 22 September 2026. Eligible trials enrolled non-clinical adults, delivered a rule-based or generative conversational agent for wellbeing, and reported wellbeing, positive affect, skill use, behaviour change or engagement. Searches covered PubMed, PsycINFO, Scopus, Web of Science, Google Scholar, CTRI, ClinicalTrials.gov and OSF preprints. Effects were Hedges g under restricted maximum likelihood (REML). Certainty was graded with GRADE. Risk of bias was assessed with RoB 2. Results. On wellbeing scales alone (three trials, N = 1,658) the pooled effect was g = 0.07 (95% CI −0.06 to 0.19). Adding the one trial that reported positive affect as the extractable endpoint (four trials, N = 2,144) raised the estimate to g = 0.17 (95% CI −0.01 to 0.35; I² = 68%; tau² = 0.022; 95% prediction interval −0.40 to 0.73). Both intervals include or approach zero. Versus wait-list or care-as-usual the effect was g = 0.28 (95% CI 0.14 to 0.42; I² = 0%). Versus active web-based wellness resources the effect was g = 0.02 (95% CI −0.10 to 0.14). Self-care behaviour in one trial was d = 0.36 at ten days and was not sustained at one month. Engagement was low: in the largest trial users chatted on a mean of 4.3 of 30 days. No completed adult non-clinical Indian trial was found. Most trials did not pre-specify or systematically report adverse events. Conclusions. Measured on wellbeing rather than symptom scales, AI conversational agents deliver a small and uncertain benefit in non-clinical adults. The benefit is visible against wait-list and not visible against good digital resources. Skill and behaviour outcomes remain sparsely measured. An Indian workplace evidence base does not yet exist. Registration. This review was not prospectively registered.