What is it about?
AI chatbots are increasingly being used for emotional support, but can they respond to a patient the way a trained therapist would? We built an AI system in which four “expert” language models, each focused on a different kind of therapeutic response (supportive, exploratory, directive, and meaning-making), suggest what to say next, and a fifth model chooses the best option. We took 300 real moments from psychodynamic therapy sessions with patients with depression and placed the AI’s responses side by side with what the human therapist actually said. Experienced clinicians, who did not know which response came from whom, rated them. The AI chose broadly similar types of responses, though not always at the same moments. Its supportive, exploratory and directive responses were rated as good as the therapists’. However, therapists were better at helping patients make deeper sense of their experiences, and overall their responses were rated as more helpful.
Featured Image
Photo by Maximilian Bungart on Unsplash
Why is it important?
Many people are already turning to AI for mental health support, yet we know little about how its responses compare to those of trained therapists in real clinical moments. Instead of asking whether AI is “good” or “bad” at therapy, our study points to where it does well and where it falls short. The AI system we tested handled supportive, exploratory and directive responses well. It struggled with meaning-making, the kind of intervention that links what a patient says in the moment to broader patterns in their life and relationships. Our findings suggest that, for now, AI is best used as a supervised aid for clinicians rather than as a replacement for them. For example, it could give therapists feedback after a session or support the training of early-career therapists. The study also shows how much clinicians and AI researchers need to work together to develop and evaluate these tools responsibly.
Perspectives
Throughout my years of studying and training as a clinical psychologist, I have felt a tension between two voices. One emphasizes the uniqueness of the human encounter in therapy, something that, in this view, cannot truly be studied, measured, or imitated. The other embraces technology, driven by a strong wish for innovation and progress. I believe that collapsing into either of these positions is risky. Dismissing technological developments means missing real opportunities to help patients and therapists. Embracing them uncritically risks losing what is most valuable in the human perspective and experience. For me, this study was a way to hold both sides: to learn closely what these models can do, while staying curious about what makes the human therapist’s response meaningful. I hope this article encourages more integrative thinking, in which clinicians and technologists learn from one another rather than talk past each other.
Yael Bar-Shachar
Bar-Ilan University
Read the Original
This page is a summary of: Evaluating large language models versus human therapists in generating interventions for patients with depression in psychodynamic psychotherapy., Journal of Consulting and Clinical Psychology, October 2026, American Psychological Association (APA),
DOI: 10.1037/ccp0001034.
You can read the full text:
Contributors
The following have contributed to this page







