Pooled study of 1,967 people finds no clear edge for AI therapy over human care

Results varied enormously between studies (I² = 98%). Acceptance of AI was similar to acceptance of human therapists but higher than for bibliotherapy. Studies using general chatbots such as ChatGPT leaned toward conventional therapy, while those using specialised tools such as Woebot leaned toward AI.

The authors rated the certainty of evidence as very low. Later and larger studies reported lower acceptance of AI, which the authors link to fading novelty or small-study effects, and the funnel plot was asymmetric, so publication bias cannot be ruled out.

The authors included studies that compared acceptance of conversational AI, such as ChatGPT and Woebot, with conventional treatment such as human therapists or bibliotherapy. Acceptance was measured on Likert-type scales covering willingness to use, attitudes, satisfaction and trust, and converted to Hedges' g so different scales could be pooled. Two researchers screened and extracted data independently, and the Kmet checklist rated all 10 studies as low risk of bias.

Subgroup analysis suggested that the type of control therapy, cultural background and AI type might explain some of the variation, while population, baseline distress and participant source did not. The cultural findings rest on single studies in some groups: Southern European and Middle Eastern samples showed different patterns, and the authors call these preliminary.

The authors note the review was not registered in a public registry, that questionnaire reliability was rarely reported, and that older AI models studied years ago may not reflect current capabilities. They conclude that AI could complement human therapists, but that the evidence does not show either approach is better.