A Supervised AI Chatbot Reduced Symptoms in a Waitlist Trial

A four-week randomized trial found that adults assigned to a purpose-built generative-AI mental-health app reported larger short-term symptom reductions than adults assigned to wait for access. The result is promising, but it belongs to a specific system and a specific safety model: clinicians and trained researchers reviewed every AI response after it had reached the participant, with expert intervention available.

Research references: [1]

What the trial tested

Researchers recruited 210 U.S. adults through Meta advertisements. Participants had to screen positive for clinically significant symptoms of depression, generalized anxiety or high risk for feeding and eating disorders; this was not a cohort established through independent clinician diagnoses. Active suicidality, mania and psychosis were exclusion criteria.

The team randomized 106 participants to Therabot and 104 to a waitlist. People in the intervention arm received daily prompts during the first four weeks and could continue using the app without prompts through week eight. The waitlist arm completed the same assessments but did not receive Therabot until the study ended.

Therabot used fine-tuned Falcon 7B and Llama 2 70B models. Its training conversations were written and reviewed by mental-health experts and drew primarily on cognitive behavioral therapy. The app also included guardrails and a crisis-classification model.

Research references: [1] [2]

Depression symptoms fell more than with waiting

Among participants who crossed the depression-screening threshold, model-estimated PHQ-9 scores fell 6.13 points with Therabot and 2.63 points with the waitlist at four weeks. At eight weeks, the declines were 7.93 and 4.22 points. The adjusted between-group effect sizes were 0.845 and 0.903, and both comparisons remained significant after correction for six primary comparisons.

The PHQ-9 result is clinically relevant because a five-point reduction is commonly treated as a meaningful within-person change. But this trial did not isolate the app’s conversational content from attention, daily prompts, expectations or novelty. A waitlist is a useful early comparator; it is not an active treatment or an equal-contact control.

Research references: [1]

Anxiety and weight-concern measures also moved

For participants meeting the anxiety-screening threshold, GAD-Q-IV scores declined 2.32 points with Therabot and 0.13 with the waitlist at four weeks. At eight weeks, the declines were 3.18 and 1.11 points. Adjusted effect sizes were 0.840 and 0.794.

Among participants classified as high risk for feeding and eating disorders, Weight Concerns Scale scores fell 9.83 points with Therabot and 1.66 with the waitlist at four weeks. At eight weeks, the declines were 10.23 and 3.70 points. Adjusted effect sizes were 0.819 and 0.627.

These are changes on self-reported screening and symptom scales. They do not establish clinician-diagnosed remission, medical recovery from an eating disorder, better daily functioning or prevention of self-harm.

Research references: [1]

Human oversight was part of the intervention

Every Therabot response was reviewed after transmission by trained clinicians and researchers. The response reached the participant first. If the system produced an inappropriate reply, such as medical advice, the team contacted the participant with a correction. If a participant raised a safety concern, staff provided guidance and emergency resources.

That design makes the oversight inseparable from the evidence. The trial does not show how the system would perform without that layer, at larger scale or among people at acute risk. It also should not be used as evidence that a general-purpose chatbot can replace a therapist or manage a crisis on its own.

The participants used the app. Of 106 people assigned to Therabot, 101 used it at least once. Users averaged 260 messages, 24 days of interaction and 6.18 hours in the app. Ninety-six completed a working-alliance questionnaire, with a mean score of 3.59 on a five-point scale.

The authors compared that score descriptively with a published outpatient-psychotherapy norm. Participants were not randomized between Therabot and human therapy, so the study cannot establish equivalent relationships, care quality or outcomes.

Research references: [1]

What the study can and cannot support

The clearest warranted conclusion is narrow: a purpose-built, expert-trained generative-AI intervention, combined with prompts and post-response human monitoring, produced larger eight-week reductions in three self-reported symptom measures than waiting among recruited adults who passed the safety exclusions.

The trial was open-label, recruited online and paid participants $25 for each of three assessments. Attrition was concentrated in the intervention arm, although overall missingness was low. The active intervention lasted four weeks and follow-up ended at week eight.

The study was funded by Dartmouth College. By July 2026, Dartmouth reported that senior author Nicholas Jacobson and first author Michael Heinz had co-founded Therabot Labs, LLC to pursue the technology’s rollout. That later commercial relationship does not invalidate the trial, but it matters when interpreting current claims about the product and should be disclosed alongside the results.

The next useful tests are active-control trials with comparable contact and expectations, longer follow-up, functional and quality-of-life outcomes, systematic reporting of model failures and adverse events, and transparent accounting of the clinical labor required to supervise the system.

Mental-health access is a real human problem, and on-demand tools may become part of the answer. This trial moves the evidence beyond anecdotes. Its responsible lesson is not that therapy has been automated; it is that any benefit observed here came from a carefully built system with humans still responsible for safety.

If you or someone you know is in immediate danger or considering self-harm, contact local emergency services or a crisis line. This article reports research and is not medical advice.

: Heinz MV, Mackin DM, Trudeau BM, et al. “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment.” *NEJM AI*. Published March 27, 2025. https://doi.org/10.1056/AIoa2400802

: ClinicalTrials.gov. “Testing an Artificial Intelligence Chatbot for Mental Health” (NCT06013137). https://clinicaltrials.gov/study/NCT06013137

: Geisel School of Medicine at Dartmouth. “Nicholas Jacobson Recognized for Foundational Therabot Trial.” Published July 16, 2026. https://geiselmed.dartmouth.edu/news/2026/nicholas-jacobson-recognized-for-foundational-therabot-trial/

Research references: [1] [3] [2]