A 210-person trial tested a generative AI therapy chatbot

What happened

On 27 March 2025, NEJM AI published 'Randomized Trial of a Generative AI Chatbot for Mental Health Treatment' by Michael Heinz, Nicholas Jacobson and eight co-authors at Dartmouth. The trial tested Therabot, described in the paper's abstract as an 'expert-fine-tuned' generative AI system built for mental health treatment, not a companion product. It enrolled 210 adults nationally, screened into three groups: major depressive disorder, generalized anxiety disorder, or high risk for a feeding or eating disorder. Participants were randomly assigned to a four-week Therabot intervention (106 people) or a waitlist control (104 people); the waitlist group received no access until after the study's eight-week endpoint.

What the documents show

A Dartmouth press release reports a 51 percent average reduction in depression symptoms, a 31 percent reduction in anxiety symptoms, and a 19 percent reduction in body-image and weight concerns, measured against the waitlist group. The abstract states primary outcomes were symptom change from baseline to the 4-week postintervention point and an 8-week follow-up, with secondary outcomes covering engagement, acceptability and therapeutic alliance, the collaborative relationship between a user and a therapeutic tool. The analysis used mixed models and effect sizes from a log-odds ratio, comparing differential change between groups rather than a simple before-and-after count.

The mechanism

A waitlist-controlled randomised trial is a different kind of evidence from the surveys and forum analyses common elsewhere in this record set. Randomising who received Therabot immediately, versus who waited, means a measured difference between groups is far less likely to be explained by who chose to seek help, or when, than in a self-selected sample. That is what lets a trial claim a treatment effect rather than a correlation. Therabot is also a distinct product category from a companion app: fine-tuned specifically as a therapeutic tool and evaluated against clinical scales, not a persona-driven companion evaluated on engagement.

What it leaves open

Dartmouth's account quotes the senior author stating that AI-powered therapy remains in critical need of clinician oversight and that no generative AI system is ready to operate fully autonomously in mental health, calling for further work to quantify the risks involved. Neither the press release nor the bibliographic record for this article states a funding source, so that cannot be confirmed here. This is an editorial reading: an eight-week trial of a purpose-built clinical tool, run by an academic team, says little on its own about consumer companion apps not built or tested to the same clinical standard.

• Was the comparison a randomised waitlist control, or a before-and-after measure with no comparison group?
• Was the chatbot purpose-built and clinically evaluated, or a general companion app used for an unintended purpose?
• Did the authors call for more oversight or further study, and does that caution travel with the result when it is cited elsewhere?

Read as intended, the trial is early but real evidence that a purpose-built, clinically supervised chatbot produced measurable symptom change against a randomised control over eight weeks. It is not evidence that a general-purpose companion app, built for a different purpose and without this trial design, would show the same result.