Eighteen months after Dartmouth researchers published the first randomized controlled trial of a fully generative AI therapy chatbot, the software at the center of that study, Therabot, is no longer just a research prototype. It is being spun into a company, tested in new clinical settings, and held up by federal regulators as a test case for how, or whether, generative AI should be allowed inside mental health care at all.
The original trial, led by Dartmouth’s Geisel School of Medicine and published March 27, 2025 in NEJM AI, remains the most rigorous evidence to date that a large language model can deliver measurable clinical benefit without a human therapist in the loop. Researchers enrolled 210 adults diagnosed with major depressive disorder, generalized anxiety disorder, or at clinically high risk for eating disorders. Of those, 106 were randomly assigned to use Therabot on their smartphones for eight weeks, split between a four-week period of unlimited access and four weeks of self-initiated use, while 104 participants in a control group got no access at all.
What the numbers actually showed
The results were striking for a field used to modest effect sizes. Participants with depression reported a 51% average reduction in symptoms. Those with generalized anxiety disorder saw a 31% average reduction, with many moving from moderate to mild anxiety or dropping below the clinical threshold for diagnosis entirely. People at risk for eating disorders showed a 19% reduction in body-image and weight concerns, outpacing the control group. Users engaged with the app for an average of six hours over the trial, roughly the equivalent of eight therapy sessions, and nearly three-quarters of participants were not receiving any other pharmaceutical or talk-therapy treatment at the time.
Michael Heinz, the trial’s first author and an attending psychiatrist at Dartmouth Hitchcock Medical Center, said participants rated their working relationship with Therabot as comparable to what patients typically report with human clinicians. Nicholas Jacobson, the study’s senior author and an associate professor of biomedical data science and psychiatry at Geisel, put it more bluntly: “We did not expect that people would almost treat the software like a friend.”
A slower road, deliberately
Jacobson has spent the past year positioning Therabot as the methodical alternative to a crowded field of consumer chatbots making therapeutic claims without controlled trials behind them. “There are a lot of folks rushing into AI for mental health,” he said in comments reported in July 2026, when he was named one of two worldwide finalists for the Chen Institute and Science Prize for AI Accelerated Research. “We took the slower road with Therabot, grounding it in evidence-based practice.” The prize ultimately went to UC Davis neuroscientist Sergey Stavisky, but the nomination put a spotlight on Jacobson’s approach: years of iterative testing, published peer review, and built-in safeguards, including automated crisis detection that prompts users in high-risk conversations to contact 911 or a suicide-prevention hotline.
Commercialization meets regulatory limbo
Jacobson has since co-founded Therabot Labs LLC to commercialize the software, and the product is now in beta testing across what the researchers describe as multiple clinical settings, with new trials wrapping up or preparing to launch. But the path to market runs through an FDA that has, as of this year, authorized zero AI-enabled medical devices specifically for mental health, despite clearing more than 1,200 AI devices in other domains like radiology and cardiology. Jacobson testified before the FDA’s Digital Health Advisory Committee in November 2025, when the panel met specifically to weigh guardrails for generative AI-enabled digital mental health devices, and he has also appeared before New Hampshire lawmakers on AI safety standards.
The skeptics are not strawmen
Not everyone in the field is convinced the Therabot results generalize. Heinz himself has cautioned that “no generative AI agent is ready to operate fully autonomously in mental health,” a notably restrained claim from a co-author of the study showing the strongest results yet. Clinician groups surveyed separately by the American Psychological Association in 2026 have echoed that caution at scale: large majorities of practicing psychologists say generative chatbots lack the nuance to safely treat mental health conditions and worry about their ability to catch warning signs a trained clinician would recognize. The 210-person Dartmouth sample, while large for an AI mental health trial, is also still small relative to the population Therabot Labs hopes to eventually serve, and critics note the eight-week trial window says little about what happens after months or years of sustained use.
What’s next
The stakes of Therabot’s commercialization attempt extend well beyond one company. If Therabot Labs secures some form of FDA clearance or finds a path to market as a wellness tool outside the medical device framework, it will become a template other generative AI mental health startups try to copy. If it stalls in regulatory review, as several comparably well-funded competitors in adjacent AI mental-health categories have in the past year, it will reinforce the view that the FDA’s current pathways are mismatched to how generative AI actually works. Either outcome will shape how millions of people who cannot access or afford traditional therapy eventually interact with AI-based mental health support, and how much clinical evidence regulators ultimately decide that interaction requires.
Photo: Open_Arms_Initiative / PIXABAY via Pixabay
