Evaluating large language models versus human therapists in generating interventions for patients with depression in psychodynamic psychotherapy

Recent advances in artificial intelligence, especially the rise of large language models (LLMs), hold great promise for transforming mental health care, but their growing integration also raises pressing questions about their ability to deliver meaningful therapeutic interventions. This study compared the capabilities of LLMs to those of human therapists in producing context-sensitive psychotherapy responses within 16-session supportive-expressive psychodynamic psychotherapy for patients with depression. Specifically, it compared LLM and human therapists on (a) the frequency and typicality of their intervention choice and (b) the quality and general helpfulness of their interventions.

We developed an LLM-based framework composed of four specialized LLM agents, each representing a different therapeutic dimension, and a Judge-LLM that selects the most appropriate intervention for a given context. A sample of 300 therapy session segments, each containing both human- and LLM-generated responses, served as the evaluation data set. Expert clinical judges assessed interventions for typicality and quality.

LLMs and human therapists relied on broadly similar intervention types. However, the contextual choice of when to use a specific type differed. Human interventions were rated as generally more helpful. Specifically, humans produced higher quality meaning-making interventions, while no differences were found for supportive, directive, or exploratory interventions.

The findings suggest that under the specific locally deployed multiagent configuration tested, LLMs can generate a range of context-sensitive therapeutic responses but showed limitations in producing high-quality meaning-making interventions. These findings highlight both the promise and the current boundaries of such systems as supervised therapeutic adjuncts. (PsycInfo Database Record (c) 2026 APA, all rights reserved).