You have probably had a chatbot pop up on a website asking what you need.
Most people click past them. A new kind of chatbot has been tested in a very different setting, and this one was built to listen to how you feel.
Researchers at Dartmouth tested a generative AI chatbot called Therabot in 210 adults with depression, generalized anxiety, or a high risk of an eating disorder. The results were published in NEJM AI in March 2025.
After eight weeks, depression symptoms fell by an average of 51 percent in the Therabot group.
Also read: Taking melatonin night after night may almost double your heart failure risk
Anxiety symptoms dropped 31 percent, and concerns about body image and weight dropped 19 percent.
It is the first randomized trial of a fully generative AI therapy chatbot in people with clinical mental health symptoms.
How Therabot was built and used
Therabot runs on a smartphone app. Users type to it and it answers in open-ended replies grounded in cognitive behavioral therapy.
That is a form of talk therapy that targets the thinking and behavior patterns driving depression and anxiety.
Also read: How to clean your home after an illness
Participants in the Therabot group had four weeks of unlimited access, followed by another four weeks in which they could still initiate conversations.
On average they used the app for about six hours across the trial, the equivalent of roughly eight therapy sessions.
They rated the sense of connection with the bot on a par with what people describe with a human therapist.
Why the authors say the tool is not ready to use on its own
The research team monitored conversations during the trial and stepped in several times to address safety concerns, including signs of suicidal thinking, and to correct responses the bot had got wrong.
Also read: Kids who sleep under 9 hours a night show brain differences that last for years
The app also prompts users to call 911 or a crisis line when it detects a high-risk message.
First author Michael Heinz, an assistant professor of psychiatry at Dartmouth's Geisel School of Medicine, was explicit about the limits.
"No generative AI agent is ready to operate fully autonomously in mental health where there is a very wide range of high-risk scenarios it might encounter," Heinz said in a Dartmouth news post.
The authors call for rigorous safety and efficacy benchmarks and continued expert supervision before any wider rollout.
Also read: 7 warning signs your cortisol may be chronically high, and what to do about it
Also read: Your gut bacteria may make your stroke harder to recover from
This article is made and published by Mie Hermansen, who may have used AI in the preparation.