Therabot: an impressive study, but no proof that therapists are redundant

Therabot: an impressive study, but no proof that therapists are redundant

11 April 2025

Last week I was pointed to a recent RCT on Therabot that generated a lot of buzz. I took the time to read the paper and wanted to share some thoughts.

Therabot is a conversational AI for mental health that was evaluated against a waitlist control for clinical symptoms of three disorders: major depressive disorder, generalised anxiety disorder, and eating disorders. Four weeks of treatment were followed by four weeks of follow-up, during which users could freely continue using the tool.

Unfortunately, not much detail is available about the actual content of the intervention. Based on the description, there didn’t appear to be an overarching structure or programme: the tool relied mainly on empathetic responses, validation, targeted interventions, and questions prompting elaboration. As for results: participants were highly engaged, averaging 260 messages and over six hours of exchanges. They also showed significant, large improvements and managed to establish a working alliance with the bot. All of this ran largely autonomously, with research staff needing to intervene only 15 times for safety concerns and 13 times to correct inappropriate responses (such as offering specific medical advice).

Does this mean therapists will soon be redundant? Not really. A human-in-the-loop proved necessary, even if interventions were thankfully rare. That kind of monitoring will likely remain both a future legal requirement and a key safeguard for quality. Additionally, participants were slightly younger and technologically minded, and open to AI. This isn’t necessarily a problem as this is probably also the group most likely to use such tools going forward. It is however too early to conclude that AI therapy is acceptable and effective for everyone. The authors themselves see Therabot more as the next iteration in digital therapeutics: not a replacement for face-to-face therapy, but a further development of internet- and mobile-based interventions.

All that said, this remains an impressive study demonstrating the promise of a custom-built conversational AI for mental health, assessed with actual patients and yielding meaningful results.

Well worth reading in Nature Mental Health.

Published
Categorized as insight