When Michael Heinz talks about the AI mental health chatbot he helped build at Dartmouth, he starts with what it isn’t. It has no photorealistic avatar, no soothing human voice, no design trick meant to make a lonely person forget they’re talking to software. “We are never optimizing toward engagement for the sake of engagement,” says Heinz, an assistant professor of psychiatry at Dartmouth and an attending psychiatrist at Dartmouth Hitchcock Medical Center. The chatbot, called Therabot, is deliberately cartoonish in appearance, a design choice meant to keep users grounded in the fact that they are interacting with a machine, not a person.
A Six-Year Project Aimed at a Widening Gap
Heinz began building Therabot in 2019 alongside clinical psychologist Nicholas Jacobson at Dartmouth’s Center for Technology and Behavioral Health. Their motivation was blunt: the United States does not have nearly enough licensed therapists to meet demand, and wait lists for outpatient mental health care routinely stretch for months. Therabot was designed to fill part of that gap not by replacing a therapist but by giving patients “a chatbot along for the ride” between sessions, walking them through cognitive behavioral therapy exercises assigned by a human clinician.
What sets Therabot apart from the wave of consumer AI companion and wellness apps is evidence. According to Heinz, it is the first generative AI mental health chatbot to show clinical results in a published trial, trained specifically on clinical data emphasizing cognitive behavioral therapy principles rather than scraped from the open internet. The tool runs on university-controlled servers and remains restricted to approved research studies rather than being available for public download.
Guarding Against the Sycophancy Problem
Heinz is candid about the risks baked into large language models used for therapeutic purposes. Commercial chatbots are frequently tuned to keep users engaged, which can push them toward being agreeable to a fault, a tendency researchers call sycophancy. In a mental health context, that can mean validating a harmful thought rather than challenging it. Heinz acknowledges that even Therabot is not immune: “there certainly were users who formed a therapeutic alliance with the bot” during clinical trials, a result he views as encouraging for engagement but one that also demands careful monitoring for unhealthy dependency.
That tension sits at the center of a fast-growing and largely unregulated market. Millions of Americans already turn to general-purpose AI chatbots for informal mental health support, according to a June 2026 American Psychological Association survey in which more than three-quarters of psychologists reported that their patients were discussing AI use in therapy sessions, from seeking a diagnosis to companionship. Unlike those consumer tools, Therabot was built from the ground up under clinical supervision, with human researchers reviewing its behavior throughout development.
A Working Psychologist Pushes Back
Not everyone in the field is convinced a chatbot, however carefully engineered, belongs anywhere near a therapy relationship. Psychologist Dominic Candido argues that round-the-clock availability risks fostering exactly the kind of dependency Heinz says he wants to avoid. Candido contends that chatbots cannot truly “read behavior or emotional reactions” or perform the kind of case conceptualization a trained clinician builds over multiple sessions. For Candido, effective therapy rests on “human empathy, which is the ability to see the world through other people’s eyes” — something he doesn’t believe software can replicate no matter how well-trained.
Candido also raises a practical concern that cuts across the entire AI mental health category: data privacy. Many consumer-facing apps are not HIPAA-compliant, he notes, and users often have no clear picture of where their most sensitive disclosures end up. “We don’t know where this information is going,” he says, a warning that applies far beyond any single product.
Why the FDA Isn’t the Answer
Heinz doesn’t expect Therabot to ever receive FDA approval, and he isn’t waiting for one. He believes the agency’s clearance process, designed around static medical devices, doesn’t fit dynamic AI systems that continue to be refined after deployment. That regulatory vacuum has pushed some states to act on their own. Vermont has already passed legislation requiring that a licensed professional review AI-generated mental health advice before it reaches a patient. New Hampshire, where Dartmouth is based, has not yet passed comparable rules, leaving researchers like Heinz to set their own internal guardrails in the meantime.
What Comes Next for Clinical AI Therapy Tools
Therabot remains confined to research settings for now, a deliberate choice that keeps it out of app stores while Dartmouth’s team gathers more trial data and refines its safety measures. Heinz’s longer-term vision is narrower than Silicon Valley’s: not a chatbot that replaces a therapist, but one that extends a clinician’s reach between appointments, always paired with human oversight rather than operating alone.
The bigger question is whether that careful, research-first model can compete with the speed of commercial AI products already reaching millions of users with far less clinical vetting. As state legislatures slowly begin drafting their own rules and psychologists like Candido continue to voice skepticism about whether empathy can be engineered, Heinz’s approach — slow, supervised, and unglamorous by design — may prove to be either the responsible template the field needs, or a cautious outlier overtaken by faster-moving rivals before it ever leaves the lab.
