Dartmouth's therapy bot cuts depression symptoms by 51%

When panic hits at 3 a.m. or depression makes getting out of bed feel impossible, getting an appointment with a therapist is usually weeks away. But can a chatbot actually make a clinical dent in mental health?

Researchers at Dartmouth College just published the results of the first-ever randomized clinical trial testing a generative AI therapy app. Published in NEJM AI, the study looked at Therabot, a conversational agent built on evidence-based psychotherapy, and found that software alone drove clinical improvements rivaling gold-standard human therapy.

Outpatient results in just six hours

The trial tracked 106 participants across the U.S. diagnosed with major depressive disorder, generalized anxiety disorder, or an eating disorder, alongside a 104-person control group. Nearly 75% of the Therabot group were not taking psychiatric medications or seeing a human therapist at the time.

Over eight weeks, users interacted with Therabot via a smartphone app. In total, participants logged an average of six hours of chat time across the study, which researchers equated to roughly eight traditional outpatient sessions.

The clinical results were striking. Participants with depression saw an average 51% symptom reduction, marking substantial real-world improvements in daily mood. Those with generalized anxiety reported an average 31% drop, with many sliding from moderate anxiety to mild or dropping below clinical thresholds altogether. Even among participants at risk for eating disorders, body-image concerns dropped 19%, significantly beating the control group.

Midnight conversations without the fear of judgment

Perhaps the most surprising takeaway was that patients formed what clinicians call a therapeutic alliance with the bot. Participants reported levels of trust, rapport, and collaboration on par with what people feel toward human clinicians.

Usage spiked during vulnerable hours like the middle of the night. Nicholas Jacobson, the study's senior author, noted that users started treating the software like a friend. A big reason was psychological safety: people felt comfortable opening up because they knew the bot wouldn't judge them.

Therabot, which Jacobson's lab has developed since 2019 alongside Dartmouth psychiatrists, relies on an original training set rooted in cognitive behavioral therapy. If it flags red-flag inputs like suicidal ideation, the app immediately surfaces buttons to dial 911 or reach crisis hotlines, while the research team stays positioned to intervene.

Why it matters

Mental healthcare has a massive supply problem. In the U.S. alone, there are roughly 1,600 patients with depression or anxiety for every licensed mental health provider. Traditional therapy is expensive, waitlists are brutal, and social stigma keeps millions from ever seeking care. For anyone stuck in that gap, a validated, conversational tool on their phone is an immediate lifeline.

We think Dartmouth's approach is smart because it treats AI not as an autonomous doctor, but as accessible infrastructure. Putting evidence-based CBT exercises into an open-ended dialogue that patients actually engage with in real time solves therapy's biggest bottleneck: consistency and accessibility.

However, the risks in this space are massive. As study co-author Michael Heinz pointed out, the open-ended nature that makes AI therapy work is also what makes it dangerous. Generative models can produce unpredictable responses in high-stakes crises. While Therabot demonstrated over 90% adherence to clinical best practices in earlier benchmarks, mental health leaves virtually zero room for error. We believe AI therapy will become an indispensable triage and maintenance tool, but letting any generative model operate in this domain without human oversight remains a non-starter.