AINEWS
Dartmouth's Therabot, seven years in, just published the first clinical trial of a generative AI mental health chatbot. Apps that share the name are already in app stores.
A patient walks out of a session with an exposure assignment, the homework that makes a therapy hour actually work, and for the first time, a chatbot is alongside them for the exercises between visits. The chatbot is called Therabot. It lives at Dartmouth's Center for Technology and Behavioral Health, and seven years into the project, the public still cannot use it.
Therabot is the first generative AI mental health chatbot to publish clinical results in a peer-reviewed trial, according to work described in NEJM AI. The work began in 2019 under psychiatrist Michael Heinz and clinical psychologist Nicholas Jacobson, before ChatGPT, and the system is still confined to approved research studies. Heinz told the Concord Monitor he could not say when, or whether, the tool will reach the market.
While Therabot sits in research, two unrelated products already ship to consumers under the same name. One is a GPT-4 tool that analyzes relationship chat logs; the other is a subscription wellness app with mood tracking. Neither is connected to Dartmouth. The name collision is not a lawsuit and the apps are not impersonating the research tool; they just beat it to the storefront.
Heinz grounds the project in an access argument. "There's a huge gap between people who need treatment and the availability for it," he told the Concord Monitor. His team is not building a chatbot that happens to be good at conversation. They are building one for the work a therapist already assigns: "You think about having this interactive intervention with you that could really go into those exposures and kind of help guide you through the different exercises," Heinz said.
General-purpose chat models are trained to be plausible. A clinical chatbot has to be right in the specific way a clinical interaction is right, which is why Heinz is blunt about the cost of getting that wrong. "We need guardrails on these models, both in training and deployment," he said.
John McCarthy and three colleagues proposed the 1956 Dartmouth Summer Research Project on Artificial Intelligence in 1955, and the proposal is the place "artificial intelligence" was first used in print. That 70-year arc is a grace note, not the spine of the story. The spine is that an institution with a long claim on the field has spent seven years trying to make a chatbot safe enough to use on real patients, and has not yet shipped one.
Vermont's H.816 would require licensed professional review of AI-generated mental health advice before it reaches a patient. A December 2025 White House executive order discourages states from regulating AI directly. If the federal preemption push holds, Vermont-style safeguards are exactly the kind of state rule that gets challenged.
For now, the asymmetry is sharp: the Therabot with a clinical trial sits behind a research consent form, and the Therabots in the app store face no such gate. Heinz will not predict when the research version ships. The next concrete milestone is whatever the NEJM AI paper details about effect size and follow-up, and whether other research groups can replicate the design.