AI Mental Health Chatbots: What the Therabot Trial Actually Shows

A product-specific study

A randomized trial published in NEJM AI in March 2025 studied Therabot, a generative system fine-tuned by mental health experts. It enrolled 210 adults with clinically significant depression, anxiety or high risk for feeding and eating disorders: 106 received four weeks of access and 104 entered a waitlist. Outcomes were assessed at weeks four and eight. This is a review of a 2025 experiment, not a newly completed study.

What improvement means

Depression scores changed by an average of −6.13 with Therabot versus −2.63 on the waitlist at four weeks, and −7.93 versus −4.22 at eight weeks. These are symptom-score changes, not percentages cured. A waitlist comparison cannot isolate effects of AI from attention, engagement and expectations. It also cannot establish equivalence or superiority to clinician-delivered psychotherapy because that comparison was not performed.

Do not generalize to every chatbot

Training, safeguards, escalation procedures and updates vary between systems. Evidence for an expert-tuned intervention cannot validate general-purpose assistants or companion apps. A software change can also separate the tested product from the version people use. Screening symptoms are not interchangeable with comprehensive diagnosis, and the eating-related group was defined by risk rather than all diagnosed eating disorders.

Safety needs its own evidence

A 210-person study cannot rule out rare serious harms. Crisis handling, privacy, inaccurate advice, harmful-belief reinforcement and overreliance require specific evaluation. An investigator-developed intervention benefits from independent replication; academic affiliation does not remove potential interests. Peer-reviewed publication is not FDA authorization. The result is promising, bounded evidence—not a recommendation to replace professionals or use chatbots for medication decisions or emergencies.

Sources

• NEJM AI primary trial report
• Dartmouth study report

Editorial disclosure: The accompanying image is a generated editorial illustration, not a photograph of study participants. This reporting is educational, not individualized medical advice.