What is the problem?
Many people with depression or anxiety cannot reach a therapist, or drop support apps after a few days. The authors see generative AI (programs that write answers in everyday language) as a way to offer personalised support at scale.
But a rigorous trial was needed to check that it really works and that people use it.
Digital therapies often struggle to keep people engaged, which is another reason to test a chatbot carefully.
What did the researchers do?
They tested Therabot, a generative AI chatbot fine-tuned by experts. The national trial recruited 210 adults with clinically significant symptoms of major depression, generalised anxiety, or at clinically high risk of eating disorders.
Participants were assigned by chance: 106 used Therabot for 4 weeks, 104 went on a waitlist and only got the app after the study ended. Symptoms were measured at 4 and then 8 weeks.
Participants were split into three groups according to screening results: depression, anxiety, or eating-disorder risk. The main outcomes were symptom changes after 4 weeks and at the 8-week follow-up. Secondary outcomes were engagement, acceptability and the therapeutic alliance. Statistical models compared the changes between groups, and effect sizes (Cohen's d, a measure of how big the gap is) were reported.
What were the results?
Therabot users saw their symptoms fall significantly more than the waitlist group. For depression, the average score fell by 6.13 points against 2.63 at 4 weeks, and by 7.93 against 4.22 at 8 weeks. For anxiety, it fell by 2.32 against 0.13 at 4 weeks.
For eating-disorder risk, the score fell by 9.83 points against 1.66 at 4 weeks. Participants used the tool a lot (over 6 hours on average) and rated the therapeutic alliance, meaning the working bond with the tool, as comparable to that with human therapists.
The gaps held at 8 weeks: depression fell by 7.93 points against 4.22, anxiety by 3.18 against 1.11, and eating-disorder risk by 10.23 against 3.70. Effect sizes were large for all three conditions, between 0.627 and 0.903.
Key figures
- 210adults randomly assigned to Therabot (106) or a waitlist (104)
- -6.13 vs -2.63average fall in depression score at 4 weeks, Therabot against waitlist
- over 6 hoursaverage use of Therabot by participants
What does this change for patients?
This trial suggests a well-built chatbot can ease symptoms alongside existing care, especially for people who cannot reach a therapist. It also suggests such a tool can stay engaging.
It does not replace a consultation. If you are going through a hard time, talk to a health professional or a helpline in your country.
The authors conclude that fine-tuned chatbots offer a feasible approach to delivering personalised mental health support at scale.
What are the limits?
The comparison group simply waited and received nothing, so the study cannot say whether Therabot beats other support or a therapist. Follow-up is short: 8 weeks in all.
The authors themselves say that research with larger clinical samples is needed to confirm effectiveness and whether the results apply more widely.
The trial is registered as NCT06013137, and the scores come from symptom questionnaires rather than from a clinician's assessment.
A chatbot is not therapy: what Mima does
Therabot is a therapeutic chatbot tested in a research trial with a specific set-up. It is not the same as a general-public app, and it does not replace a professional.
Mima is not therapy. She is a listening app, without judgement, that keeps a record of what you live through so you can talk about it with your team. If you are struggling, speak to a professional. In an emergency, call an emergency number.
Frequently asked questions
Can a chatbot treat depression?
In this 4-week trial of 210 adults, symptoms fell more with Therabot than with a waitlist. This does not prove it can replace a therapist, and larger studies are needed.
What is Therabot?
It is a generative AI chatbot fine-tuned by experts, tested to support symptoms of depression, anxiety and risk of eating disorders.
What does a waitlist comparison group mean?
Those participants did not get the app during the study and got it afterwards. This compares against no treatment, but not against another treatment.
Can I already use Therabot?
The paper describes a research trial and does not present Therabot as an available treatment. If you are in distress, contact a health professional.
The original study
Heinz MV, Mackin DM, Trudeau BM, Jacobson NC, et al. “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment.” NEJM AI, 27 March 2025. DOI : 10.1056/AIoa2400802
Read the full study in NEJM AIThis article summarises a scientific study and is not medical advice.