AI in Mental Health Enters the Clinic: With Study Results, Why Can It Still Not Claim to Replace Therapists?

Known Facts: The Study Subjects Are Not All Chatbots

Dartmouth published an introduction to the Therabot trial in March 2025, reporting an intervention group and a control group that did not receive Therabot, and stating that researchers evaluated the conversations. The institution also emphasized that generative AI mental health applications still require clinical supervision. This article is based on the institution's public introduction and does not purport to have reviewed all methods and supplementary materials of the original paper.

Research Judgment: Look First at the Comparator, Then at the Effect Size

Comparison with people who did not receive the tool answers the question of the effect of adding a tool under specific study conditions. It does not directly answer whether the tool is superior to therapists, usual care, other chatbots, or low-cost self-help materials. These questions require different study designs.

At procurement, the included and excluded populations, follow-up duration, attrition, actual human involvement, model version, and adverse events should be recorded separately. A single proportion of symptom change cannot substitute for the full body of evidence.

Product Decisions: Bind Evidence to Specific Uses

For campus or employee support products, a more reasonable starting point is to clearly state the goal: health education, mood tracking, service navigation, or treatment adjunct. Each use requires corresponding risk controls and acceptance metrics. After expanding the audience, changing the model, or altering human support, whether the original evidence still holds must be reassessed.

Editorial Judgment: Possible Implications

• Establish linked records of product version—applicable population—use scenario—evidence.

• Efficacy claims in marketing copy should return to the original study conditions, and evidence should not be borrowed across products.

What to Watch Next

• Whether replication studies with active comparators emerge.

• Longer follow-up, independent team validation, and real-world adverse event disclosure.

Limitations and Boundaries of Use

• This article is not treatment advice, nor does it constitute a clinical recommendation for any product.

• No original patient data review was conducted; no best AI therapy product was rated on this basis.

Original Source Identifiers

[therabot] [who-guidance]