Dartmouth Trial: AI Chatbot Reduces Depression Symptoms by 51%
A first-of-its-kind study suggests generative AI can provide mental health benefits comparable to traditional outpatient therapy.
Researchers at Dartmouth have completed the first clinical trial of Therabot, a generative AI-powered therapy chatbot, demonstrating that AI-assisted care can produce clinically meaningful improvements in mental health. The study, published March 27 in NEJM AI, indicates that the software can effectively reduce symptoms of depression, anxiety, and eating disorders over a short period.
The eight-week trial involved 106 participants diagnosed with generalized anxiety disorder, major depressive disorder, or an eating disorder. According to the study results, participants saw an average reduction in depression symptoms of 51% and a 31% decrease in generalized anxiety symptoms. Additionally, those at risk for eating disorders reported a 19% reduction in concerns regarding weight and body image. Nicholas Jacobson, an associate professor of psychiatry and biomedical data science, noted that these improvements were comparable to those typically reported in traditional outpatient therapy.
Addressing the Provider Gap
The development of Therabot began in 2019 at Dartmouth's AI and Mental Health Lab. The system was built using a training set rooted in cognitive behavioral therapy (CBT) best practices and evidence-based psychotherapy. This technical foundation was designed to address a critical systemic failure in the U.S. healthcare system: a severe shortage of mental health professionals. Currently, the U.S. averages only one provider for every 1,600 patients suffering from anxiety or depression, leaving millions without immediate access to care.
The Human-AI Alliance
Beyond the raw data, the trial revealed an unexpected psychological shift in how patients interacted with the software. Jacobson observed that participants appeared to form genuine relationships with the bot, treating the software almost like a friend. This suggests that generative AI can establish a "therapeutic alliance"—the trust and collaboration between patient and provider—that was previously thought to be the exclusive domain of human clinicians. By scaling this support, AI could potentially bridge the treatment gap for populations that lack the financial or geographic means to access a human therapist.
Safety and Future Implementation
Despite the positive outcomes, researchers warn against the immediate, unsupervised deployment of such tools. Michael Heinz, an assistant professor of psychiatry, emphasized that no generative AI agent is currently ready to operate fully autonomously. The primary concern remains the wide range of high-risk scenarios that a mental health patient might encounter, which require human judgment and intervention to ensure safety. Moving forward, the industry will likely view software-based therapy as a complement to, rather than a replacement for, in-person clinical care, provided that rigorous safety benchmarks are established and maintained.