Therabot

Development began in 2019, before ChatGPT arrived. According to the team’s account, two early attempts failed: a model trained on online peer-support forums reproduced depressive talk instead of treating it, and a model trained on psychotherapy transcripts produced generic, even stereotyped, responses. The team then decided to write therapist-patient dialogues itself, reviewed by a psychiatrist and a clinical psychologist and based mainly on CBT. It reports more than 100,000 hours of human work.

In the version evaluated in 2024, these dialogues were used to fine-tune two open language models (Falcon-7B and LLaMA-2-70B), hosted on Amazon Web Services servers. A classifier flags high-risk messages, such as suicidal ideation, and displays an emergency module that points to helplines.

The trial, published in NEJM AI in March 2025 (Heinz et al., 2025), enrolled 210 adults recruited in the United States through Facebook and Instagram ads. All screened positive, on self-report questionnaires, for clinical symptoms of major depression, generalized anxiety, or a high risk of eating disorder. Half received the app for four weeks, with a daily prompt to use it, then free access for four more weeks. The other half was placed on a waitlist.

These effect sizes measure the difference between the two groups. Engagement was high: 95% of participants used the app, for six hours on average over the study, with 260 messages on average. Participants also completed a therapeutic alliance questionnaire, whose mean scores the authors present as comparable to those of outpatient psychotherapy patients.

The trial measures therapeutic alliance with the WAI-SR, a questionnaire designed for the relationship between a patient and a human therapist. Mean scores (3.6 out of 5) are presented as comparable to outpatient norms, without a statistical test. In detail, bond and agreement on goals are slightly below these norms, agreement on tasks slightly above. Asked how similar it was to a “real therapist,” participants gave an average of 4.9 out of 7.

In 2026, Dartmouth describes Therabot as in “beta testing,” with several trials in different clinical settings. A single-arm trial in 15 people with cannabis use disorder ended in April 2026, with no results published to date. An NIH grant awarded in September 2026 funds a feasibility study comparing Therabot with a version enhanced by passive sensing (sleep, activity) and biofeedback.

Jacobson and Heinz co-founded Therabot Labs, LLC to roll out the tool “in accordance with clinical standards.” Asked in August 2026, Michael Heinz could not say when, or whether, Therabot would reach the market. Nicholas Jacobson says the tool is intended for clinical populations, to be expanded “in a judicious way, with good clinical oversight”; Michael Heinz says the team never optimizes engagement for its own sake.

Therabot belongs on this site not because a clinician in France could use it (they cannot), but because it has become a textbook case. It is often cited as proof that a generative chatbot “heals.” The trial is real and serious, and its effects are large (1). It was also run by the tool’s designers, against a waitlist, with clinicians reviewing every response.

It is this gap between the result and the way it circulates that concerns clinicians. One simple question helps measure it: what exactly does this study compare, and who was watching the tool while it ran? Asked of Therabot, it takes nothing away from the trial’s value. Asked of apps that can already be downloaded, which may invoke results of this kind without having produced any themselves, it is very revealing.

Finally, Therabot shows a trajectory: an academic project becomes a company, and the current version, with its voice and avatar, is already no longer the one that was evaluated. Before recommending a tool of this kind to a patient, it is worth checking which version was studied, by whom, and against what.