The useful question about an AI mental-health companion is whether interacting with it improves symptoms. Dartmouth’s Therabot moved that question into a controlled experiment: people could use a smartphone chatbot when difficulties arose, with a research team checking its conversations. Results published on 27 March 2025 showed greater symptom improvement among people offered the app.[2]
The study randomly assigned 210 US adults with clinical-level depression or anxiety symptoms, or elevated eating-disorder risk: 106 received Therabot and 104 joined a waitlist. Daily prompts encouraged conversations for four weeks. Access continued for four more weeks without those prompts; the comparison group received access after the study. Consequently, the eight-week assessment measured a period when users could still talk to the chatbot.[3]
The measured changes
Symptom changes at eight weeks
| Outcome and scale | Therabot group | Waitlist group |
|---|---|---|
| Depression: Patient Health Questionnaire-9, 0–27 | −7.93 points; 70 participants | −4.22 points; 69 participants |
| Anxiety: Generalized Anxiety Disorder Questionnaire-IV, 0–13 | −3.18 points; 52 participants | −1.11 points; 56 participants |
| Weight Concerns Scale, 0–100 | −10.23 points; 42 participants | −3.70 points; 47 participants |
All three questionnaire outcomes favoured Therabot at four and eight weeks after adjustment for multiple comparisons. A waitlist comparison measures the benefit of offering the intervention; a trial against active treatment would answer the separate question of comparative clinical effectiveness.[1]
A conversation built around therapeutic methods
Therabot was developed from original dialogues based on established psychotherapy practices, including cognitive behavioural therapy. It produced open-ended replies and personalised its questions as conversations progressed. Users could respond to a scheduled check-in or start a conversation themselves, including during the night. That flexibility let the software engage with a problem at the moment someone wanted to discuss it.[2]
The trial registration describes training conversations written by the research team and an intervention intended to deliver supportive behavioural techniques. This is a specific treatment design: the text examples help shape how the system responds to distress, asks questions and suggests ways of working through a problem. A general-purpose chatbot’s ability to sound supportive does not establish that it reproduces this intervention.[4]
Engagement was substantial: Dartmouth’s research centre reports that 95% of the intervention group interacted with Therabot, with users averaging about six hours of use. Participants also rated trust, shared goals and collaboration favourably. Those ratings describe their experience of a therapeutic relationship; they do not turn the waitlist experiment into a comparison of clinical effectiveness against human therapists.[3]
The human work inside the trial
The study team intervened 15 times for safety concerns and made 13 corrections after inappropriate responses. These were interventions after messages had been sent, so the monitored service included the capacity to contact participants and respond to problems. The figures describe events, not necessarily 28 different people. Removing that clinical support would change what had been tested.[3][1]
The registered eligibility criteria excluded people currently being treated for active suicidality, bipolar I or II disorder, or active psychosis. That defines the population the investigators set out to recruit. Findings from this selected adult sample cannot resolve effectiveness across every mental-health condition or during a crisis.[4]
The next question is how to use the service
In July 2026, Dartmouth reported that Therabot remained in beta testing, with further studies in different clinical settings approaching completion or preparing to launch. The researchers had also formed Therabot Labs to develop a service with clinical monitoring of safety and outcomes. That update describes a development programme, rather than a completed demonstration of a universally available therapy.[5]
Dartmouth’s research centre identifies an active comparison and time-comparable therapy as the next research steps.[3]
Such a comparison would help separate several possibilities: benefit from the therapeutic content, benefit from having a responsive conversation readily available, and benefit from attention and encouragement. The answer would shape where a service belongs in care: supporting people between appointments, reaching those unable to access therapy, or providing a defined intervention of its own. Each role needs evidence that the service can reliably fulfil it.
Sources & further reading
- Randomized Trial of a Generative AI Chatbot for Mental Health Treatment NEJM AI · 27 Mar 2025 · Accessed 11 Oct 2026
- First Therapy Chatbot Trial Yields Mental Health Benefits Dartmouth · 27 Mar 2025 · Accessed 11 Oct 2026
- Randomized Trial of a Generative AI Chatbot for Mental Health Treatment: Research Brief Dartmouth Center for Technology and Behavioral Health · Accessed 11 Oct 2026
- Testing an Artificial Intelligence Chatbot for Mental Health — NCT06013137 ClinicalTrials.gov · Accessed 11 Oct 2026
- Nicholas Jacobson Recognized for Foundational Therabot Trial Dartmouth Technology Transfer Office · 16 Jul 2026 · Accessed 11 Oct 2026