The first clinical trial of a generative AI therapy tool has shown that it might actually help people manage depression, anxiety, and eating disorder risks. That’s a huge step forward for the use of artificial intelligence in mental health—but experts warn that it doesn’t mean the flood of AI therapy apps currently on the market are safe or effective.

The study, published in NEJM AI (a journal under the New England Journal of Medicine), tested a generative AI chatbot called Therabot, developed by a team of researchers at Dartmouth’s Geisel School of Medicine. And while its results were impressive—comparable to traditional talk therapy—it also revealed the challenges of applying such tools responsibly and ethically.

The Breakthrough

Therabot wasn’t just another chatbot spitting out generic wellness phrases. It was carefully trained on evidence-based therapy methods, not just scraped internet conversations or Reddit posts. The developers knew that conventional generative AI models could go dangerously off-script—especially when dealing with vulnerable individuals.

“We tried training the model on conversations from forums, but it ended up sounding more like a caricature of therapy,” explained Nick Jacobson, lead researcher and associate professor at Dartmouth. “So we created custom datasets based on actual therapeutic techniques.”

The result? An AI model that delivered structured, supportive responses rooted in real therapy practices like cognitive behavioral therapy (CBT).

The Trial

In an eight-week clinical trial involving 210 people, half were given access to Therabot, while the other half were not. These individuals were dealing with depression, anxiety, or were at high risk of developing eating disorders.

Participants using Therabot reported major improvements:

• Depression symptoms dropped by 51%

• Anxiety symptoms dropped by 31%

• Concerns about body image and weight fell by 19%

Most users engaged regularly with the bot—sending an average of 10 messages per day—mirroring levels of interaction usually seen in human therapy sessions. And here’s the surprising part: The progress made in eight weeks with Therabot was similar to what patients typically achieve with 16 hours of traditional therapy.

Why It Matters?

Only about half of people with mental health conditions receive treatment, and those who do may get just one short session a week. In that context, Therabot offers something radical—support that’s available 24/7, instantly responsive, and potentially scalable.

But that’s also where things get tricky.

Don’t Trust Every “AI Therapist” Just Yet

Despite the trial’s success, Jacobson is cautious. He says the findings should not be seen as a blanket endorsement of the countless AI therapy tools flooding the market.

“Most of them aren’t built on evidence-based practices,” he said. “And they don’t have trained researchers monitoring what the bots are saying. That’s a big concern.”

During the trial, Jacobson personally reviewed messages to ensure the AI didn’t say anything inappropriate or harmful. But in the real world, these AI therapy apps are mostly unsupervised. And unlike Therabot, many are just variations of large language models like Meta’s Llama or OpenAI’s GPT, which can unknowingly reinforce harmful thoughts—especially around sensitive topics like eating disorders or suicidal ideation.

If an AI bot encourages someone to lose weight without understanding their medical history, the consequences can be devastating.

The Regulatory Gap

Jacobson emphasized that when AI bots are marketed as legitimate therapeutic tools, they fall under the regulatory scope of the Food and Drug Administration (FDA). But so far, enforcement has been lacking.

“My suspicion is that none of the major therapy bots out there today would pass FDA clearance if they actually had to prove their claims,” he said.

That regulatory vacuum could expose vulnerable users to harm—and it raises serious ethical concerns about how these tools are being positioned and used.

The Risk of Misuse

Jean-Christophe Bélisle-Pipon, a professor of health ethics at Simon Fraser University, echoed similar concerns. He says the trial is promising, but warns that without proper oversight and integration into official healthcare systems, these tools could backfire.

If people don’t have access to approved AI therapy tools, they might start using general-purpose bots like ChatGPT or Character.AI for emotional support. That’s risky, especially as some bots have been caught engaging in problematic or even sexually explicit conversations with users, including minors.

The Bottom Line

Therabot is a glimpse of what responsible, research-backed AI therapy could look like. It’s a major step forward in making mental health care more accessible—but it’s also a wake-up call for tech companies, regulators, and users.

This study doesn’t mean AI is ready to replace therapists. But it does show that, with the right design, oversight, and ethical considerations, generative AI can support mental health care in powerful ways.

Until then, the promise of AI therapy should be met with cautious optimism—and a lot more scrutiny.

ALSO READ: Ethics in AI: Balancing Innovation with Responsibility