Randomized Trial of a Generative AI Chatbot for Mental Health Treatment

Key Findings

• — National RCT, N = 210 adults (Therabot n = 106, waitlist n = 104), stratified by screening result into major depressive disorder, generalized anxiety disorder, or clinically high risk for feeding and eating disorders; the waitlist received app access after eight weeks.
• — Primary outcomes were symptom change from baseline to four weeks and to eight-week follow-up; secondary outcomes were engagement, acceptability and therapeutic alliance, analysed with cumulative-link mixed models and log-odds-based effect sizes.
• — Per the university's summary of the published results, the Therabot group showed an average 51% reduction in depressive symptoms, a 31% reduction in anxiety symptoms, and a 19% reduction in body-image and weight concerns, each outpacing control.
• — The same summary reports that about 75% of the Therabot group were not receiving pharmaceutical or other therapeutic treatment at the time, and that participants rated the therapeutic alliance comparably to human therapists in prior literature.
• — Therabot was fine-tuned on expert-written therapeutic dialogue rather than scraped counselling data, and the trial ran with clinician monitoring of conversations.

Methodology Notes

Preregistered randomized controlled trial with a waitlist control (no active comparator), four-week intervention and eight-week follow-up, US adults recruited nationally, self-report symptom scales; published 2025-03-27 in NEJM AI (DOI 10.1056/AIoa2400802). The journal page is behind a challenge wall from this workspace and PubMed does not index the article, so the record was verified via Crossref (title, authors, date) and the Semantic Scholar deposited abstract, which carries the background and methods; the outcome percentages are taken from Dartmouth's 2025-03-27 news release rather than the article text and should be re-checked against the paper. A well-known 2025 result that had never been entered in the library.

Sources