Generative AI chatbot effectively reduces symptoms of depression, anxiety, and eating disorders in clinical trial

A groundbreaking randomised controlled trial conducted by Dartmouth College has found that a generative artificial intelligence (GenAI) chatbot can significantly reduce symptoms of depression, generalised anxiety, and eating disorders. The study is the first of its kind to assess the clinical efficacy of a GenAI-based mental health intervention, offering a promising step forward in the use of artificial intelligence for therapeutic purposes.

The research, led by clinicians at Dartmouth, evaluated Therabot, an AI-powered mental health chatbot developed in 2019. In total, 210 participants were enrolled in the study, of whom 106 presented with clinically significant symptoms of major depressive disorder, generalised anxiety disorder, or were identified as being at high risk for feeding and eating disorders. These participants received access to Therabot and were encouraged to interact with the chatbot daily for 30 days. Following this intervention phase, participants could continue to use the chatbot for an additional four weeks, though without daily reminders. Assessments were conducted at baseline, four weeks, and eight weeks.

Therabot offers structured mental health support through a conversational interface. Users may either initiate sessions on demand or respond to scheduled notifications. The chatbot delivers tailored psychological interventions through question prompts, empathetic replies, and affirming statements designed to meet the user’s expressed needs. According to the developers, this continuity of interaction fosters a deeper therapeutic experience, closely mirroring aspects of traditional human-delivered psychotherapy.

Despite the growing popularity of GenAI applications in wellness and companionship, they have not previously been studied for their capacity to deliver clinical-grade mental health care. “To date, conversational agents using GenAI have fallen under general purpose, wellness, or companion applications, rather than software intended for the diagnosis and treatment of mental health disorders,” the authors noted in the study.

The results were striking. Participants diagnosed with major depressive disorder experienced a 51% reduction in symptoms at both the four- and eight-week assessments. Those living with generalised anxiety disorder saw a 31% reduction in symptoms, while individuals at risk for feeding and eating disorders experienced a 19% reduction. Notably, none of the participants were taking psychiatric medication during the study period, suggesting that the improvements were attributable to the chatbot intervention itself.

Engagement levels were high: participants used Therabot on average 24 days out of the 30-day intervention period and spent approximately 6.18 hours interacting with the chatbot. On average, participants sent 260 messages, indicating sustained and active use.

The study also explored users’ perceptions of the chatbot’s utility and emotional connection. Participants reported that Therabot was intuitive and easy to use, and many said they felt better after engaging with it. Moreover, the quality of the therapeutic relationship—referred to as therapeutic alliance—was found to be comparable to that experienced with human therapists. The authors observed that key elements of psychosocial therapy, including empathy and shared goals, were effectively established through Therabot—something previously considered difficult or impossible to achieve with non-generative AI systems.

As part of the safety protocol, clinicians closely monitored Therabot’s responses. In 15 instances, they contacted participants due to safety concerns such as suicidal ideation. On 13 occasions, researchers intervened to correct inappropriate or inaccurate chatbot responses. Despite these incidents, the authors concluded that the intervention was overall safe and well tolerated.

Importantly, the study emphasised the broader implications of scalable AI-powered mental health tools. While traditional, evidence-based psychosocial treatments are known to be effective, they are often labour-intensive and limited by resource constraints. As the study notes, “Although empirically validated psychosocial treatments exist, they are resource intensive, and limited in scalability and accessibility, leading to fewer than half of the people with a mental health disorder receiving care.”

The potential for GenAI tools to alleviate pressure on overstretched behavioural health systems is significant, especially as waiting lists for mental health services continue to grow. Unlike rule-based AI chatbots, which rely on predefined responses set by programmers, GenAI systems offer dynamic and personalised interactions. This adaptability enables a deeper and more human-like therapeutic experience, which may enhance both efficacy and engagement.

If further validated through larger trials and real-world implementation studies, tools like Therabot could transform the digital mental health landscape, making quality support more accessible to individuals across a wide range of conditions.

Participants in the control arm of the study did not receive access to Therabot until after the conclusion of the trial.