Therabot showed promise for mental health—but the human safety net mattered
Therabot（THER-uh-bot）
A chatbot specially designed and trained for mental health support.
waitlist control
A comparison group that waits before receiving the tested tool.
active control
A comparison treatment that gives people another real activity or service.
What happened
On March 27, 2025, researchers from Dartmouth College published a study in NEJM AI. It tested Therabot, a generative AI chatbot built for mental health support. The original study and a Dartmouth research brief describe the trial.
The study included 210 adults with clinically significant symptoms. Participants screened for major depressive disorder, generalized anxiety disorder, or high risk for feeding and eating disorders. Researchers randomly assigned 106 people to Therabot. They assigned 104 people to a waitlist control group.
How the trial worked
The Therabot group received daily prompts for four weeks. They could use the chatbot during that period. They then kept access for another four weeks without daily prompts. The waitlist group did not get app access during the study. It received access after the study ended.
Researchers measured symptoms at four and eight weeks. The Therabot group showed greater reductions across all three symptom areas. This is important evidence because the study used random assignment. It still measures symptom scores, however. It does not prove that Therabot cures a disorder or makes a diagnosis.
Why this was a special chatbot
Therabot was not ordinary ChatGPT used without changes. Experts created and reviewed dialogue examples for the system. The examples mainly used cognitive behavioral therapy. The chatbot also included safety tools, such as a system for detecting possible crises.
This design changes how we should read the result. The trial tested a specialized product. It also tested a research process with safety monitoring. The result does not automatically apply to every general-purpose chatbot.
What users reported
About 95% of people assigned to Therabot used it. That was 101 participants. Users sent an average of 260 messages. They used the system on an average of 24 days. Total use averaged 6.18 hours during the study period.
Users also reported a strong working relationship with the chatbot. The average Working Alliance score was 3.59. The research brief says this was similar to scores seen in traditional outpatient psychotherapy. This was a user rating, not proof of equal treatment quality.
Safety still required people
Safety monitoring was not just a background detail. Staff intervened 15 times because of safety concerns. Examples included expressions of suicidal ideation. Staff also corrected 13 inappropriate responses, including medical advice from the chatbot.
Those numbers do not show that the system cannot help. They show that human oversight mattered in this trial. A fully independent chatbot was not what the study tested.
What remains unknown
The comparison group was a waitlist. It did not use another chatbot, receive a similar digital activity, or meet with a therapist. Therefore, the study cannot fully separate Therabot's specific treatment effect from hope, attention, or natural change over time.
The follow-up ended at eight weeks. We do not know how long the improvement lasts. The trial also excluded people with active suicidality, mania, or psychosis. Its findings therefore do not cover every mental health situation.
What to watch next
The researchers say future studies should use an active control. They also recommend comparing Therabot with time-matched in-person therapy. Longer follow-up and broader participants will matter too. For now, the fairest conclusion is cautious: a carefully designed mental health chatbot may reduce symptoms for some adults. This trial does not show that AI can replace human care.
A mental health chatbot helped in one careful study
📰 Full story: Therabot showed promise for mental health—but the human safety net mattered
A special chatbot reduced several symptoms in a study of 210 adults. People still needed to watch its replies.
Therabot（THER-uh-bot）
A chatbot made for talking about mental health.
generative AI
AI that creates a new reply from what someone writes.
waitlist control
A group that waits before using the tested tool.
💡 The gist
- Therabot users had fewer depression, anxiety, and eating concerns.
- Researchers studied 210 adults for eight weeks.
- Human staff still handled safety problems.
Therabot is a mental health chatbot made by Dartmouth College researchers. It uses generative AI. That means it writes new replies from a person’s message. It was designed for mental health support.
The researchers ran a randomized controlled trial. In this kind of study, people are placed into groups by chance. One group had 106 adults. They used Therabot for four weeks. The app sent daily reminders during that time. They could keep using it for four more weeks without reminders.
The other group had 104 adults. It was a waitlist control group. Those people did not use the app during the study. They could use it after the study ended. This gave researchers a comparison group.
At weeks four and eight, everyone answered questions. The Therabot group showed larger drops in three areas. These were depression symptoms, anxiety symptoms, and eating concerns. A symptom is a sign that someone is struggling. The results do not prove that the chatbot cures a mental illness.
Therabot was not regular ChatGPT. Experts wrote and checked many example conversations. The system used ideas from cognitive behavioral therapy. This therapy helps people notice thinking and behavior patterns. The chatbot also had tools for finding possible crises.
Most people in the Therabot group used the system. The number was 101 people, or 95 percent. They sent an average of 260 messages. They used it for about 24 days. Total use averaged 6.18 hours.
People also said the chatbot felt helpful and easy to use. Their ratings suggested a working relationship with Therabot. Those ratings were similar to some ratings for human therapy. They did not prove that Therabot worked just like a therapist.
Safety is an important part of the story. Staff stepped in 15 times because of safety concerns. They corrected 13 chatbot replies that were not suitable. Some replies gave medical advice. This means the chatbot was not working alone.
The study used a waitlist, not another active treatment. Hope and attention may have helped too. Researchers still need to compare Therabot with another chatbot and with human therapy. They also need longer studies. This study shows promise under careful research rules. It does not show that AI can replace a therapist or help every person.
Can a talking computer help a sad or worried mind?
📰 Full story: Therabot showed promise for mental health—but the human safety net mattered
A special talking computer helped some adults feel less upset.
Therabot（THER-uh-bot）
A talking computer made for hard feelings.
AI
A computer that reads words and makes replies.
Dartmouth College
A school that studies health and computers.
Dartmouth College is a school that studies health and computers.
It tested Therabot, a special AI for hard feelings. Therabot reads a message and writes a reply. It is not a human.
The study included 210 adults. 106 adults used Therabot for four weeks. 104 adults waited. Later, the waiting group could use it too.
At four and eight weeks, people answered questions. Therabot users had less sadness and worry. They also had fewer eating troubles.
People watched the computer’s replies. They helped 15 times when safety worries appeared. They fixed 13 replies that were not right.
So the computer was not alone. People with the biggest danger did not join the study. The study did not compare Therabot with a real therapist.
This was one study. It says Therabot may help some adults. It does not say every talking computer helps.