In the summer of 1956, a small group of researchers spent eight weeks at Dartmouth College pursuing an audacious premise: that human intelligence could be described precisely enough for a machine to simulate it.
The proposal that convened them, written the year before by Dartmouth mathematician John McCarthy and three colleagues, had given the subject a name — the first time the phrase “artificial intelligence” appeared in print. The workshop is now widely regarded as the birth of the field.
Seventy years later, Dartmouth psychiatrist Michael Heinz is testing one of the harder claims within that premise: that a machine can do therapy. He describes a patient who leaves a session with an assignment and, instead of facing it alone, has a chatbot along for the ride.
“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,” said Heinz, the first author on the clinical trial behind Therabot.
Therabot, built inside Dartmouth’s Center for Technology and Behavioral Health, is the first generative AI mental health chatbot to show clinical results in a published trial. Heinz, an attending psychiatrist at Dartmouth Hitchcock Medical Center, began the work in 2019 with clinical psychologist Nicholas Jacobson — four years before ChatGPT made the technology a household name.
Seven years in, the Dartmouth Therabot isn’t available to the public. It exists only inside approved research studies, and Heinz couldn’t say when, or whether, it will reach the market. Apps calling themselves AI therapists, meanwhile, are already downloadable — reaching users years ahead of trial data.
Other products have even used the same name. A GPT-4 tool that analyzes uploaded chat logs for relationship problems calls itself Therabot; so does a subscription wellness app with mood tracking. Neither is connected to Dartmouth.
Heinz thinks his tool could fill a gap and do it ethically. “There’s a huge gap between people who need treatment and the availability for it,” he said. A chatbot doesn’t keep office hours or have a waitlist.
He is also blunt that the same qualities make it risky. “With mental health, we thought about all the guardrails,” he said. “We need guardrails on these models, both in training and deployment.”
The distinction he keeps returning to is between tools designed for this work and tools that merely happen to be good at conversation: “There’s a big difference between somebody interacting with Therabot and somebody interacting with ChatGPT for mental health symptoms.”
Many questions, fewer answers
Dominic Candido has spent 35 years as a psychologist, and the always-available quality Heinz describes is exactly what gives him pause.
“That’s one benefit that’s touted of the apps, is access, where people might not have access otherwise,” Candido said. “But it’s unfettered access, too. 24 hours a day.”
In his practice, he is not available that way, on purpose. “I’m available for emergencies, but not to ask a question,” he said. “Somehow I think that sets up, can set up a dependency on this tool to be there when they want them.”
Whether constant availability is the point of these tools or the flaw in them sits underneath many of the arguments unfolding around AI and mental health care: in Dartmouth’s labs, in statehouses, and in the practices of clinicians whose patients arrive having already talked things through with a machine.
Candido founded Hanover Psychiatry for the Geisel School of Medicine and spent eight years on its psychiatry faculty. He now leads Enhance Health, a private practice in the Upper Valley, and has no involvement in the Therabot project. He credits it for being trained on clinical data and built on cognitive behavioral therapy — “terrific,” he said. With most apps, by contrast, “there’s no way of knowing if there’s any kind of modality that’s being conveyed.” His questions are about what therapy requires that a chatbot can’t supply.
But he is still wary of the technology because, good therapy, in his view, starts before technique — with what he calls conceptualization: figuring out who a patient actually is before deciding what to do for them. Someone might arrive carrying panic disorder, social anxiety, recurrent depression and an avoidant personality at once, and he explained it requires hearing more than what a patient says out loud.
“The app can’t read behavior or emotional reactions. It’s only responding to language that’s being input,” Candido said. “There’s a total person here that is expressing emotion, is acting in certain ways that I’ve learned to read over the years.”
He teaches his trainees to listen for what a patient is thinking, not just what they’re saying — exactly what he says a chatbot can’t do. “The app can hear what you’re saying, but can’t listen to what you’re thinking, and that’s the important part.”
Increasingly, that work is being attempted by software.
How can a chatbot be a therapist?
A chatbot like ChatGPT or Google’s Gemini is built by feeding a program enormous amounts of text scraped from the internet until it becomes extremely good at predicting what word comes next — good enough that its responses read as fluent, reasonable, even empathetic. The biggest of these systems are called “frontier” models: general-purpose, closed to public inspection, and licensed out to other apps that rent access without much control over what happens inside.
“These general frontier models are incredibly — many of them incredibly useful. I use a lot of them for different things myself,” Heinz said. “But they’re general purpose. None of these was built for mental health, for that purpose.”
Dartmouth took a different approach. Heinz said the team trained its own model from scratch, on data written to match therapeutic best practices, and hosts it on servers the university controls. “We are never optimizing toward engagement for the sake of engagement,” he said. “It really is optimized toward measures of human wellbeing, human mental health, human flourishing — that is sort of our north star.”
That thinking extends to what the tool looks like. Some apps marketed as AI therapists greet users with photorealistic human avatars. Therabot’s is deliberately cartoonish, closer to a Pixar sidekick than a person — warm enough that someone in distress will talk to it, Heinz said, but never so convincing they lose track of what they’re talking to. “This is a machine,” he said. “This is a machine that’s been trained to do a certain task.” Asked whether it’s a real therapist, Therabot answers: “No, I’m not. I’m not a human therapist. I am an AI agent that’s been trained by researchers at Dartmouth College.”
Heinz doesn’t claim the design solves the problem. Asked whether people might form real attachments to a friendly-looking robot, he said they might: “Certainly there’s risk of that.”
The pull toward agreeableness is there from the start: “At the foundation, these models — we’ve seen they can be sycophantic, they can be overly agreeable, they can be sort of engaging for the sake of engagement.” It showed up in the trial.
“There certainly were users who formed a therapeutic alliance with the bot, and I think were disappointed for the trial to end,” he said. His team builds a formal hand-off into the end of each study. He doesn’t treat over-dependence as a machine-only failure, either. “Even in real therapy, this comes up.”
What we think
Public feeling is mixed. In a University of New Hampshire survey conducted in May, nearly two-thirds of Granite Staters said they expect AI to harm the country over the next decade, and 65 percent said it would damage personal relationships. Yet two-thirds also use AI themselves — mostly for web searches and general questions. Only 4 percent said they used it for therapy or companionship.
Clinicians live in the same contradiction. In the American Psychological Association’s 2025 Practitioner Pulse Survey, the share of psychologists who had never used AI at work fell from 71 percent to 44 percent in a single year. Their worries grew alongside that adoption: two-thirds cited potential data breaches, and roughly six in 10 flagged unanticipated social harms, biased outputs, a lack of rigorous testing, and inaccurate results.
Heinz says the data question is part of why Dartmouth built its own model rather than renting someone else’s. “We use HIPAA standards for data protection,” he said. Only approved researchers see study conversations, and his team has never used them for training — that would require consent from every participant, and stripping out identifying details is harder than it sounds. “It’s not enough to just remove names,” he said. “You could look at those data … and likely identify who people are.”
The same survey helps explain why some are reaching for these tools. Nearly half of psychologists reported no openings for new patients, and 40 percent were keeping a waitlist. More than a third accept no health insurance at all — so for a patient who can’t pay out of pocket, the number of available therapists is smaller still.
What a clinician provides
What no app carries, in his account, is responsibility for the person on the other end. “I have fiduciary responsibility to do whatever is in the best interests of the patient,” Candido said. “The app doesn’t really have that.”
Candido said he has logged more than 12,000 hours of suicide interventions, and says AI is useful for research and surfacing information, “but it never supplements what clinical decisions really need to be.” Nor, he said, does it possess “human empathy, which is the ability to see the world through other people’s eyes. I have to see the world through their eyes, the person I’m working with, not my own.”
Privacy across the industry remains wide open, in his view, whatever Dartmouth does. “We don’t know where this information is going,” he said. “It’s often not HIPAA compliant, so the data itself is not protected. That’s a big hole right now.”
He said he hasn’t recommended these tools to patients, and asks the ones using them to inform him, because “we might be working at cross purposes.”
A debate about regulation
No federal agency has settled who oversees any of this. Heinz doesn’t think Therabot will ever win Food and Drug Administration approval — not because of any flaw in it, but because tools built to keep changing don’t fit a system designed for things that hold still. “The FDA is more used to looking at things that are static, that are fixed in time — drugs, apps that aren’t going to change,” he said. “It may be kind of trying to fit a square peg into a round hole.” Candido wants the agency involved anyway: “I don’t know why this is totally unregulated. It doesn’t make sense to me.”
States have moved into that vacuum, against an executive order discouraging them from regulating AI directly. Vermont and New Hampshire took the warning differently.
Vermont passed a law. H.816, signed in June, requires a licensed professional to review AI-generated mental health advice before it reaches a patient. Rep. Daisy Berbeco introduced it after “stories of people doing self-harm … when they interact with chatbots” put the risk on her radar, she told the Valley News.
Federal pressure didn’t worry her. “I don’t think Vermont’s attorney general has ever shown any sort of fear of anything that our White House has sent forward that does not support health and wellness of Vermonters,” she said. The law’s scope is narrow by design: “This isn’t a tech regulation bill … it’s not to regulate companies.”
New Hampshire hasn’t passed anything so far. State Sen. Howard Pearl, a Loudon Republican, tried regulating AI mental health tools as unlicensed practice — a state can set licensing rules even if it can’t regulate the technology directly. He called it “perhaps a workaround” for the federal order. Stakeholders couldn’t agree, and the bill was folded into a study committee.
Pearl, who owns and operates a farm, offered his own metaphor: “When the tractor came along and replaced the horse, the farmer was still driving. I think we’ve always got to make sure the farmer is still driving here.”