Large Language Models Acting as Psychotherapists: A Scoping Review
Authors: Yinzhou Wang, Ningman Gao, Yimeng Wang, Bianca Augustine, Yixuan Zhang
PsyArXiv Preprint ·
Read Preprint (opens in a new tab)Overview
As large language models increasingly enter mental health applications, understanding their therapeutic design and evaluation becomes essential.
This scoping review examines existing research on LLMs acting as psychotherapists through a human–computer interaction perspective. It investigates how these systems are designed, adapted, and evaluated, and identifies important methodological and ethical challenges.
Research Methodology
The study reviews 76 publications using the NIH Stage Model as an organizing framework.
It examines three major dimensions:
- Psychotherapeutic characteristics, including therapeutic orientation, intervention structure, and target populations.
- Technical approaches, including model adaptation methods and datasets.
- Evaluation practices, including effectiveness, usability, safety, privacy, and other assessment dimensions.
The review combines qualitative thematic synthesis with quantitative summaries of the research landscape.
Key Findings
Therapeutic Approaches
Many existing systems focus on cognitive behavioral therapy or do not explicitly identify a therapeutic orientation. Interventions are frequently limited to single-session interactions and general well-being goals.
Technical Development
Systems commonly employ prompting and fine-tuning approaches, sometimes incorporating retrieval or sensing. The review also examines 23 datasets, highlighting differences between synthetic datasets and real-world therapeutic conversations.
Evaluation and Safety
Evaluation methods vary considerably across studies, frequently relying on custom measures or LLM-based evaluation. Safety, privacy, and diversity receive comparatively limited attention, highlighting the need for more standardized and human-centered assessment.
My Contribution
I contributed throughout the research process, including literature review, data collection and extraction, data processing, and thematic coding. I helped categorize and synthesize findings across studies through qualitative and quantitative analysis, and contributed to the interpretation of results and manuscript writing.
Publication
Yinzhou Wang, Ningman Gao, Yimeng Wang, Bianca Augustine, and Yixuan Zhang. (2025). Large Language Models Acting as Psychotherapists: A Scoping Review. PsyArXiv.
DOI: https://doi.org/10.31234/osf.io/8x974_v1 (opens in a new tab)