The Care Gap
A Dartmouth chatbot trial, its authors' cautions, and the shortage behind both
Researchers report symptom improvements in a trial of an AI therapy chatbot, and say it still needs clinician oversight. Behind both: nowhere near enough providers.
In a study run out of Dartmouth, 106 people from across the United States talked to a chatbot through a smartphone app. They typed out responses to prompts about how they were feeling, or started conversations themselves when they needed to talk. Each of them had been diagnosed with major depressive disorder, generalized anxiety disorder or an eating disorder.
The chatbot is called Therabot. It has been in development in Nicholas Jacobson's AI and Mental Health Lab at Dartmouth since 2019. The researchers describe the study, with results published March 27, 2025 in NEJM AI, as the first-ever clinical trial of a generative-AI-powered therapy chatbot. They report significant improvements in participants' symptoms.
Jacobson, the study's senior author, explains why that matters in one sentence: “There is no replacement for in-person care, but there are nowhere near enough providers to go around.” For every available provider in the United States, he says, there is an average of 1,600 patients with depression or anxiety alone. That is his figure, as Dartmouth reported it.
There is a wider gap, too, and it is harder to see. The World Health Organization's Commission on Social Connection has published a 213-page report finding that “social isolation and loneliness are widespread, with serious but under-recognized impacts on health, well-being, and society.” The report calls on “policy-makers, researchers and all sectors to treat social health with the same urgency as physical and mental health.”
Here is what the Dartmouth researchers report. People diagnosed with depression saw a 51% average reduction in symptoms, which the researchers say led to clinically significant improvements in mood and overall well-being. Participants with generalized anxiety reported an average reduction in symptoms of 31%. Among those at risk for eating disorders, Therabot users showed a 19% average reduction in concerns about body image and weight. The researchers say that result significantly outpaced a control group.
That control group was 104 people with the same diagnosed conditions who had no access to Therabot. One more detail shapes how to read the results: almost 75% of the Therabot group were not under pharmaceutical or other therapeutic treatment at the time. These numbers come from Dartmouth's own account of the study. The full paper is in NEJM AI.
A federal advisory committee took up the same questions. The FDA's Digital Health Advisory Committee advises the Commissioner of Food and Drugs on issues, policies and regulations related to digital health technologies. On November 6, 2025, the committee discussed and made recommendations on generative-AI-enabled digital mental health medical devices. According to the FDA, it discussed the benefits, the risks to health, and the risk mitigations that might be considered for these devices, including premarket evidence and postmarket monitoring. In plain terms, that means proof before a product is sold, and watching it after.
The people who built Therabot are careful about what their results mean. The researchers conclude that AI-powered therapy “is still in critical need of clinician oversight,” even though it has the potential to provide “real-time support for the many people who lack regular or immediate access to a mental-health professional.” Michael Heinz, the study's first author, said in Dartmouth's account that no generative AI agent is “ready to operate fully autonomously in mental health where there is a very wide range of high-risk scenarios it might encounter.”
The app has one fixed response for the most serious moments. If Therabot detects high-risk content, such as suicidal ideation, it shows a prompt to call 911 or contact a suicide prevention or crisis hotline, with the press of an onscreen button. Heinz names the trade-off at the center of the technology himself: “the feature that allows AI to be so effective is also what confers its risk—patients can say anything to it, and it can say anything back.”
If you or someone near you is in crisis, contact your local emergency services. In the U.S., call or text 988 to reach the 988 Suicide & Crisis Lifeline. Elsewhere, findahelpline.com lists local services.
“There is no replacement for in-person care, but there are nowhere near enough providers to go around.”
