HugoScore hugoscore.org

Full review

Therabot CAIHL draft report

Evidence-linked HugoScore draft report for a health AI tool that affects patients.

Back to profile Markdown source

HugoScore CAIHL Draft Report: Therabot

Status: Draft for human review Last reviewed: 2026-06-10 Review method: Public-source review of the NEJM AI trial publication record, Dartmouth press releases from 2025 and 2026, ClinicalTrials.gov registration, and trial-listing pages; no contact with the research team, no app access, and no review of unpublished follow-up data. Service: Therabot Vendor: Dartmouth College (Jacobson AI and Mental Health Lab, Center for Technology and Behavioral Health), with spinout Therabot Labs, LLC Category: Mental health AI

Summary

Therabot is a generative AI therapy chatbot developed at Dartmouth since 2019 under clinician guidance. In March 2025 it became the first generative AI chatbot tested in a randomized controlled trial for mental health treatment, published in NEJM AI. The trial enrolled 210 US adults with major depressive disorder, generalized anxiety disorder, or high risk for feeding and eating disorders. Therabot users showed a 51 percent average reduction in depression symptoms, 31 percent in anxiety, and 19 percent in eating-disorder concerns against a waitlist control, with therapeutic alliance ratings comparable to human therapists. Human clinicians monitored conversations throughout and were ready to intervene. As of this review Therabot remains research-only. No public product was found. A Therabot-CALM trial for cannabis use disorder with anxiety or depression is recruiting, and the inventors have co-founded Therabot Labs, LLC, working with Dartmouth's technology transfer office toward licensing.

From a CAIHL perspective, Therabot matters less as a tool patients can use today and more as a proof of what responsible development looks like in this category. It was built by clinicians over years, trained on purpose-written therapy dialogue rather than scraped internet data, gated behind trial enrollment, monitored by humans, and evaluated with a registered randomized trial using standardized clinical measures. That is the evidence standard the marketed consumer chatbots have not met. The open question is whether those safeguards survive the transition from lab to company.

Evidence Reviewed

CAIHL Profile

  • Who does this AI serve? Research participants and the evidence base. There is no advertising, no data monetization, and no institutional purchaser. The authors' spinout introduces future commercial interests worth tracking.
  • Can patients tell AI is involved? Yes. Participants enroll through informed consent and know they are talking to an AI with human monitoring behind it.
  • Can patients meaningfully choose? Partial. Enrollment is voluntary with withdrawal rights, but the public cannot choose Therabot at all. Access is trial-gated by design.
  • Can patients correct or challenge what the AI produces? Partial. The study team reviewed conversations and could intervene when responses fell outside best practices, and IRB oversight applies. No patient-facing correction workflow exists because no public product exists.
  • Does it help patients understand or act? Yes, under trial conditions. Symptom reductions were clinically significant, engagement was high, and high-risk content triggered prompts to call 911 or crisis lines.

Agency Interpretation

Therabot's clearest agency contribution is to the whole category rather than to individual users today. It demonstrated that a generative therapy chatbot can be built carefully, supervised by humans, and tested with the same rigor expected of a clinical treatment, and that under those conditions it can produce real symptom improvement and genuine therapeutic alliance. That gives patients and evaluators a benchmark for judging every marketed mental health chatbot: if a research lab can run a randomized controlled trial, a funded company can too.

The unresolved tension is the path from trial to market. The trial's safety came from features that do not scale cheaply: clinician review of conversations, a team equipped to intervene immediately, and tight enrollment criteria. Therabot Labs, LLC now exists to license the technology, and the study authors are its founders, which is both how academic translation normally works and a conflict of interest for future evidence. Whether a commercial Therabot keeps human oversight, trial-grade evaluation, and honest scope claims will determine its eventual CAIHL posture. Until then the honest classification is research-stage and unclear in deployment.

Key Unknowns

  • Whether any public or commercial release is planned, and on what timeline. None was found as of 2026-06-10.
  • Whether the eight-week trial benefits persist longer term, and how Therabot performs against an active control rather than a waitlist.
  • How Therabot Labs, LLC will structure human oversight, data governance, and pricing in any licensed product.
  • Results and design details of the recruiting Therabot-CALM trial and any other follow-up studies.
  • Whether the team intends to pursue FDA authorization for a treatment claim.
  • What data practices governed trial conversations beyond standard research consent, including retention and reuse for model development.

Publication Recommendation

Ready for human review as a draft profile, with the status field kept at research stage, not publicly available. Revisit when Therabot Labs announces a product, when follow-up trial results publish, or if availability changes. Confidence should not rise above medium until deployment-stage governance is observable.