Can a Therapy Bot Help? What the Therabot Study Really Showed

Here is the short answer. The Therabot study, a randomized trial published in NEJM AI in March 2025, found that adults who used a purpose-built AI therapy bot for four weeks saw depression symptoms fall by an average of 51%, anxiety by 31%, and eating-disorder concerns by 19%, compared with people on a waitlist. It is the strongest evidence so far that a chatbot can deliver something like therapy.

It is also the most misquoted. Headlines turned “a carefully engineered bot helped adults under close supervision” into “AI therapy works.” Those are different claims, and the gap between them matters a lot if you are the parent of a teenager who talks to ChatGPT at night. This is an honest read of what the trial showed, how Therabot was built, and why it is not the app on your kid’s phone.

Key Takeaways

• Therabot was tested on 210 adults with depression, generalized anxiety or high risk for eating disorders. No children or teens were included.
• The comparison group was a waitlist, people who got nothing during the study, which makes a positive result easier to reach than a comparison with human therapy.
• The bot was built on expert-written therapy dialogues grounded in cognitive behavioral therapy, not on internet chat, and researchers reviewed its messages during the trial.
• The researchers themselves say no generative AI agent is ready to work on its own in mental health.
• The useful lesson for parents is a checklist: who built a mental health app, what trials it has, who watches it, and what it does in a crisis.

What the Therabot study is

Therabot is a generative AI therapy chatbot developed since 2019 in Nicholas Jacobson’s AI and Mental Health Lab at Dartmouth’s Geisel School of Medicine, with input from psychologists and psychiatrists. The trial, “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment,” was led by Michael Heinz and published on March 27, 2025.

The design, from the paper’s abstract:

• Who: 210 US adults with clinically significant symptoms of major depressive disorder, generalized anxiety disorder, or at clinically high risk for feeding and eating disorders.
• Groups: 106 got four weeks of access to Therabot. 104 were placed on a waitlist and got access only after the study ended.
• Measures: Standard symptom questionnaires at four weeks and again at eight weeks, plus engagement, acceptability and therapeutic alliance (how much users felt a trusting, collaborative bond).

According to Dartmouth’s announcement, nearly 75% of the Therabot group were not receiving medication or other therapy at the time, which is exactly the population that turns to apps.

What the Therabot trial found

The results were large by the standards of mental health research. Compared with the waitlist, Therabot users showed significantly greater drops in symptoms at both four and eight weeks:

• Depression: average symptom reduction of 51%. The paper reports effect sizes (Cohen’s d) of 0.85 to 0.90, which researchers usually call large.
• Anxiety: average reduction of 31%, with many people moving from moderate to mild anxiety.
• Eating-disorder risk: a 19% average reduction in body image and weight concerns.

Engagement was high. Users averaged more than six hours with the bot over the trial, which Dartmouth compares to about eight therapy sessions, and they often started conversations themselves, including late at night. Participants rated their therapeutic alliance with Therabot as comparable to what people report with human therapists.

Jacobson told MIT Technology Review that the gains resembled those from about 16 hours of human-delivered therapy, reached in about half the time. That comparison is across different studies, not head to head, so treat it as a rough benchmark.

How Therabot was built, and why that is the whole story

The interesting part, for an engineer, is what did not work first.

In 2019 the team trained a model on mental health conversations from internet forums. Jacobson said it produced therapy clichés: “Really tropes of what psychotherapy would be, rather than actually what we’d want.” Next they tried thousands of hours of real therapy transcripts, and that also fell short. What finally went into the trial model was a custom dataset the team wrote themselves, based on evidence-based practices, mostly cognitive behavioral therapy.

Why does that matter? A language model reproduces the patterns in its training data. Train it on internet chat and you get internet chat: agreeable, generic, sometimes harmful. Train it on carefully written examples of what a good therapist says, including when to challenge a distorted thought, and you shift what it predicts. That is the opposite of how most consumer chatbots are tuned, which is largely toward answers users rate highly.

Then there was supervision. The Dartmouth team reviewed conversations to check the bot was following best practice, and Jacobson personally reviewed incoming messages early in the trial. When high-risk content like suicidal thinking appeared, the app offered a button to call 911 or a crisis line, and the team was ready to step in. Heinz said that happened infrequently. Jacobson pointed out to MIT Technology Review that this kind of oversight would limit how many people a wider rollout could reach.

The researchers’ own caveats are clear. Heinz: “no generative AI agent is ready to operate fully autonomously in mental health.” Jacobson: “There is no replacement for in-person care.”

If you want the basics of why chatbots behave this way, our explainer on how large language models like ChatGPT work walks through it.

What the Therabot study did not show

A fair reading lists the limits, and there are several:

1. Adults only. Every participant was an adult. Adolescents are not small adults. Their brains, their risks and the way they relate to technology differ, and the APA warns that young people “may place too much trust in AI.”
2. A waitlist is a weak comparison. Beating “nothing” is the first step. It does not show the bot is as good as a human therapist, and it cannot separate the bot’s content from the effect of simply having something to talk to every day.
3. Short follow-up. Outcomes were measured at eight weeks. We do not know whether gains lasted, faded or carried hidden costs like dependence.
4. Self-reported symptoms. Questionnaires are standard tools, but they are what people say about themselves.
5. Heavy human supervision. The safety of the trial depended partly on researchers reading conversations, which is not how apps work at scale.
6. One product, one team. One trial by the developers is a promising start, not a settled result. Independent replication is what turns promise into evidence.

Jean-Christophe Bélisle-Pipon of Simon Fraser University, who was not involved, summed it up for MIT Technology Review: “We remain far from a ‘greenlight’ for widespread clinical deployment.”

Therabot versus the chatbot your kid actually uses

The APA’s November 2025 health advisory draws the same line. It says the research pointing to possible benefits does “not include general purpose GenAI chatbots,” and that even for fit-for-purpose tools there is “a lack of high-quality, large-scale clinical trials.” Here is how the evidence stacks up.

Question	Therabot (in the trial)	General chatbot (ChatGPT, Gemini, Character.AI)
Built for mental health care?	Yes, with clinicians since 2019	No, built for general tasks
Training data	Expert-written, CBT-based dialogues	Mostly internet text, tuned to user ratings
Tested in a randomized trial for treatment?	Yes, one trial in adults	No such trial for this use, per the APA
Tested in teens?	No	No treatment trials
Human review of conversations	Yes, during the trial	Only limited safety review in some products
Crisis handling	Crisis button and a team ready to intervene	Varies; Stanford and Common Sense found failures
Available to your family today?	No, it was a research tool	Yes, free on any phone

That last row is the reason this matters. Your teenager is not using Therabot. If they use AI for feelings, they are using a general chatbot, which a Stanford team found could show stigma and respond unsafely to simulated crises. A nationally representative survey in JAMA Pediatrics found that 19.2% of US 12- to 21-year-olds had used chatbots for mental health advice in 2025, and 63.3% told no one, per RAND.

How to use this evidence as a parent

Use the Therabot recipe as a checklist for any mental health app

The trial tells you what a trustworthy tool looks like. Ask four questions about any app your teen uses for feelings: Who built it, and were clinicians involved? Has it been tested in a published trial, and in kids or adults? Does a human ever review risky conversations? What exactly happens if your child types something frightening? If the app cannot answer, treat it as a general chatbot.

Separate “it felt helpful” from “it helped”

In the RAND survey, 91.7% of young users rated chatbot advice helpful, and the researchers warn that may reflect chatbots’ tendency to flatter. Therabot’s results came from symptom scales compared against a control group. When your teen says a bot helps, believe that it feels good, and then ask whether they are sleeping better, seeing friends more, or worrying less.

Keep a person in the loop

The single strongest safety feature in the Therabot trial was humans watching. At home, the equivalent is that someone knows. If your teen sees a therapist, tell the therapist about any app use; the APA recommends this so providers can catch advice that clashes with treatment. Our guide to how CBT works for kids explains what real treatment should look like.

What not to do: treat one good trial as a green light

It is tempting to read “51% reduction” and download the nearest AI therapy app. Most apps on the store have no trial at all, and the one that does was supervised by researchers. If your child is struggling, start with their pediatrician or a licensed counselor.

What to Watch For Over the Next 3 Months

• Week 4: Know which apps, if any, your teen uses when stressed, and whether any claims clinical backing. Check those claims against a published study.
• Month 2 red flags: Long nightly chats, hiding app use, pulling away from friends, or saying the bot “gets” them better than people do. The APA lists these kinds of changes as reasons to talk to a professional.
• Month 3 self-check: Has anyone outside the family, such as a doctor or counselor, heard about your teen’s app use? Watch for new research too: independent replications of Therabot-style trials, especially any in adolescents, would change this picture.

Frequently Asked Questions

Did the Therabot study prove AI therapy works?

It showed that one carefully built, supervised bot reduced symptoms in adults more than being on a waitlist did, over eight weeks. That is a real result. It did not show that AI matches human therapy, that it is safe without oversight, or that it works for teens.

Can my teenager use Therabot?

Not as a consumer app. Therabot was a research tool tested in a clinical trial with adults. Any app claiming to be “like Therabot” should be able to point to its own published trial and explain who reviews risky conversations.

Why can’t ChatGPT just do what Therabot does?

Training and tuning. Therabot learned from expert-written therapy dialogues, and humans watched its replies. General chatbots learn mostly from internet text and are tuned toward answers people like, which can mean agreeing too readily. The APA says the benefit studies do not cover general chatbots.

Were there any safety problems in the trial?

The researchers report that interventions for acute safety concerns or inappropriate replies were infrequent, and the app offered crisis-line access when risk appeared. But the trial had humans reviewing messages, so its safety record does not transfer to unsupervised apps.

What should I do if my teen is depressed and won’t see a therapist?

Start with their pediatrician, who can screen and refer. Keep talking without pressure. If you see warning signs of suicide, call or text 988 in the US, text HOME to 741741, or find a line at findahelpline.com.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Heinz, M. V., Mackin, D. M., Trudeau, B. M., et al., & Jacobson, N. C. (2025). “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment.” NEJM AI. https://doi.org/10.1056/AIoa2400802
2. Dartmouth College. (2025, March 27). “First Therapy Chatbot Trial Yields Mental Health Benefits.” Dartmouth News. https://home.dartmouth.edu/news/2025/03/first-therapy-chatbot-trial-yields-mental-health-benefits
3. MIT Technology Review. (2025, March 28). “The first trial of generative AI therapy shows it might help with depression.” MIT Technology Review. https://www.technologyreview.com/2025/03/28/1114001/the-first-trial-of-generative-ai-therapy-shows-it-might-help-with-depression/
4. American Psychological Association. (2025, November). “APA Health Advisory on the Use of Generative AI Chatbots and Wellness Applications for Mental Health.” APA. https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-chatbots-wellness-apps-mental-health.pdf
5. Moore, J., et al., & Haber, N. (2025). “Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers.” ACM FAccT 2025 (arXiv preprint). https://arxiv.org/abs/2504.18412
6. McBain, R. K., Cantor, J. H., et al. (2026). “AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults.” JAMA Pediatrics. https://doi.org/10.1001/jamapediatrics.2026.2015
7. RAND. (2026, June 1). “Nearly 1 in 5 U.S. Adolescents and Young Adults Use AI Chatbots for Mental Health Advice.” RAND Corporation. https://www.rand.org/news/press/2026/06/nearly-1-in-5-us-adolescents-and-young-adults-use-ai.html