Can AI Actually Work as Therapy? One Trial Cut Depression Symptoms by 51%

Health Science Hub Editorial

Evidence review & editorial synthesis

Nearly half of severely distressed generative-AI users in a new study said they had already turned to AI for emotional support, coping advice or mental-health help. That raises an obvious question: does talking to an AI actually make people feel better?

Increasingly, the answer appears to be yes. In the first randomized trial of a fully generative AI therapy chatbot, people using a system called Therabot experienced large reductions in symptoms. Depressive symptoms fell by about 51% on average, anxiety symptoms by 31%, and eating-disorder-related concerns by 19%.

Perhaps even more surprising, users reported something researchers once assumed would be difficult for software to reproduce: a strong therapeutic relationship.

A therapy chatbot produced unusually large improvements

Dartmouth researchers tested Therabot in 210 adults with clinically significant symptoms of major depression, generalized anxiety or high risk for an eating disorder. Participants were randomly assigned either four weeks of access to Therabot or to a waitlist control group.

Therabot was not simply a general chatbot instructed to act supportive. It had been specifically developed for mental-health treatment and fine-tuned using conversations based on established psychotherapy practices, including cognitive behavioral therapy.

After treatment, the Therabot group improved significantly more than the control group across all three symptom groups. Dartmouth summarized the change from baseline in especially intuitive terms: depressive symptoms dropped by an average of 51%, anxiety symptoms by 31%, and concerns involving body image and weight by 19%.

The underlying randomized comparison was also substantial. Effect sizes favoring Therabot were roughly 0.85 to 0.90 for depression, 0.79 to 0.84 for anxiety and 0.63 to 0.82 for eating-disorder-related symptoms across the study assessments.

Users did not merely try the chatbot once and forget about it. They spent more than six hours interacting with Therabot on average during the trial.

People appeared to form a bond with the AI

One of the trial's most intriguing findings had little to do with symptom scores.

Psychotherapy normally depends partly on what researchers call the therapeutic alliance: the patient's sense that the therapist understands them, is working with them and can be trusted. Therabot participants completed established measures of this relationship and rated their alliance with the chatbot at levels comparable to those commonly reported for human therapists.

Researchers also noticed that users frequently initiated conversations themselves, including at times when they were struggling. The system was available whenever those moments occurred rather than only during a scheduled appointment.

That does not mean people literally mistook Therabot for a person. It does suggest that some of the psychological ingredients of a therapeutic relationship, such as disclosure, perceived understanding, collaboration and repeated interaction, can emerge during a conversation with software.

For AI-mediated mental-health care, that could be an important capability. A technically competent intervention is of little use if people do not want to talk to it.

Therabot is no longer an isolated result

A 2026 systematic review collected 39 randomized trials of mental-health chatbots. Across 38 trials involving 7,401 participants, chatbots significantly reduced depressive symptoms. Across 34 trials involving 7,621 participants, they also reduced anxiety.

The average effects across this much broader literature were smaller than those seen with Therabot: Hedges' g was 0.31 for depression and 0.28 for anxiety. Interestingly, depressive-symptom effects were larger in clinical and subclinical populations than among people without clinically meaningful symptoms.

A separate review of commercial mental-health chatbots similarly found an improvement in depression, although the average benefit was modest and the certainty of evidence was low to very low.

So Therabot should not be treated as proof that every chatbot is an effective therapist. Its unusually large effect came from a system deliberately designed for mental-health treatment and tested under research supervision.

People are already using ordinary AI this way

That distinction matters because the real world is moving faster than the clinical evidence.

In the newly published study of 584 adult generative-AI users in Israel, 29.3% said they had already used systems such as general-purpose GenAI for emotional support, coping advice or mental-health assistance. Among participants reporting severe psychological distress, the proportion reached 47%.

The study cannot tell us whether those conversations helped. But it shows that informal AI-mediated mental-health care is not a hypothetical future behavior. People experiencing substantial distress are already doing it.

The strongest evidence of benefit currently comes from purpose-built systems such as Therabot rather than ordinary consumer AI, and Therabot has not yet been tested head to head against human psychotherapy.

Still, the scientific question has shifted. It is becoming harder to ask simply whether a machine can provide psychologically useful conversation. Randomized trials suggest that it can. The more consequential question is how far that capability can ultimately go.

Scientific References & Further Reading


Disclaimer: This article is for informational purposes only and does not constitute medical advice. It is not a substitute for diagnosis, treatment, or individualized medical care.