���϶ij�

���϶ij� researchers conducted the first-ever clinical trial of a generative AI-powered therapy chatbot and found that the software resulted in significant improvements in participants’ symptoms, according to results .

People in the study also reported they could trust and communicate with the system, known as Therabot, to a degree that is comparable to working with a mental health professional.

The trial consisted of 106 people from across the United States diagnosed with major depressive disorder, generalized anxiety disorder, or an eating disorder. Participants interacted with Therabot through a smartphone app by typing out responses to prompts about how they were feeling or initiating conversations when they needed to talk.

People diagnosed with depression experienced a 51% average reduction in symptoms, leading to clinically significant improvements in mood and overall well-being, the researchers report. Participants with generalized anxiety reported an average reduction in symptoms of 31%, with many shifting from moderate to mild anxiety, or from mild anxiety to below the clinical threshold for diagnosis.

Among those at risk for eating disorders—who are traditionally more challenging to treat—Therabot users showed a 19% average reduction in concerns about body image and weight, which significantly outpaced a control group that also was part of the trial.

Remote video URL

This demonstration video shows Therabot responding to a user with generalized anxiety with open-ended dialog developed from therapy best practices. (Video courtesy of Michael Heinz and Nicholas Jacobson) 

The researchers conclude that while AI-powered therapy is still in critical need of clinician oversight, it has the potential to provide real-time support for the many people who lack regular or immediate access to a mental-health professional.

“The improvements in symptoms we observed were comparable to what is reported for traditional outpatient therapy, suggesting this AI-assisted approach may offer clinically meaningful benefits,” says, the study’s senior author and an associate professor of biomedical data science and psychiatry at the .

“There is no replacement for in-person care, but there are nowhere near enough providers to go around,” Jacobson says. For every available provider in the United States, there’s an average of 1,600 patients with depression or anxiety alone, he says.

“We would like to see generative AI help provide mental health support to the huge number of people outside the in-person care system. I see the potential for person-to-person and software-based therapy to work together,” says Jacobson, who is the director of the treatment development and evaluation core at ���϶ij�’s .

, the study’s first author and an assistant professor of psychiatry at CTBH and Geisel, says the trial results also underscore the critical work ahead before generative AI can be used to treat people safely and effectively.

“While these results are very promising, 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,” says Heinz, who also is an attending psychiatrist at . “We still need to better understand and quantify the risks associated with generative AI used in mental health contexts.”

Therabot has been in development in Jacobson’s at ���϶ij� since 2019, and included continuous consultation with psychologists and psychiatrists affiliated with ���϶ij� and ���϶ij� Health.

Image

Therabot users, all previously diagnosed with a mental health disorder, experienced significant improvements in symptoms after eight weeks. (Graphic by LaDarius Dennison) 

When people initiate a conversation with the app, Therabot answers with natural, open-ended text dialog based on an original training set the researchers developed from current, evidence-based best practices for psychotherapy and cognitive behavioral therapy, Heinz says.

For example, if a person with anxiety tells Therabot they have been feeling very nervous and overwhelmed lately, it might respond, “Let’s take a step back and ask why you feel that way.” If Therabot detects high-risk content such as suicidal ideation during a conversation with a user, it will provide a prompt to call 911, or contact a suicide prevention or crisis hotline, with the press of an onscreen button.

The clinical trial provided the participants randomly selected to use Therabot with four weeks of unlimited access. The researchers also tracked the control group of 104 people with the same diagnosed conditions who had no access to Therabot. 

Almost 75% of the Therabot group were not under pharmaceutical or other therapeutic treatment at the time. The app asked about people’s well-being, personalizing its questions and responses based on what it learned during its conversations with participants. The researchers evaluated conversations to ensure that the software was responding within best therapeutic practices.

After four weeks, the researchers gauged a person’s progress through standardized questionnaires clinicians use to detect and monitor each condition. The team did a second assessment after another four weeks when participants could initiate conversations with Therabot but no longer received prompts.

After eight weeks, all participants using Therabot experienced a marked reduction in symptoms that exceed what clinicians consider statistically significant, Jacobson says.

These differences represent robust, real-world improvements that patients would likely notice in their daily lives, Jacobson says. Users engaged with Therabot for an average of six hours throughout the trial, or the equivalent of about eight therapy sessions, he says.

“Our results are comparable to what we would see for people with access to gold-standard cognitive therapy with outpatient providers,” Jacobson says. “We’re talking about potentially giving people the equivalent of the best treatment you can get in the care system over shorter periods of time.”

Critically, people reported a degree of “therapeutic alliance” in line with what patients report for in-person providers, the study found. Therapeutic alliance relates to the level of trust and collaboration between a patient and their caregiver and is considered essential to successful therapy.

One indication of this bond is that people not only provided detailed responses to Therabot’s prompts—they frequently initiated conversations, Jacobson says. Interactions with the software also showed upticks at times associated with unwellness, such as in the middle of the night.

“We did not expect that people would almost treat the software like a friend. It says to me that they were actually forming relationships with Therabot,” Jacobson says. “My sense is that people also felt comfortable talking to a bot because it won’t judge them.”

The Therabot trial shows that generative AI has the potential to increase a patient’s engagement and, importantly, continued use of the software, Heinz says.

“Therabot is not limited to an office and can go anywhere a patient goes. It was available around the clock for challenges that arose in daily life and could walk users through strategies to handle them in real time,” Heinz says. “But the feature that allows AI to be so effective is also what confers its risk—patients can say anything to it, and it can say anything back.”

The development and clinical testing of these systems need to have rigorous benchmarks for safety, efficacy, and the tone of engagement, and need to include the close supervision and involvement of mental-health experts, Heinz says.

“This trial brought into focus that the study team has to be equipped to intervene—possibly right away—if a patient expresses an acute safety concern such as suicidal ideation, or if the software responds in a way that is not in line with best practices,” he says. “Thankfully, we did not see this often with Therabot, but that is always a risk with generative AI, and our study team was ready.”

In evaluations of earlier versions of Therabot more than two years ago, more than 90% of responses were consistent with therapeutic best-practices, Jacobson says. That gave the team the confidence to move forward with the clinical trial.

“There are a lot of folks rushing into this space since the release of ChatGPT, and it’s easy to put out a proof of concept that looks great at first glance, but the safety and efficacy is not well established,” Jacobson says. “This is one of those cases where diligent oversight is needed, and providing that really sets us apart in this space.”