Leveraging Large Language Models for Simulated Psychotherapy Client Interactions: Development and Usability Study of...

Area of Science:

• Artificial Intelligence
• Psychology
• Clinical Training

Background:

• Large language models (LLMs) demonstrate advanced human-like text generation capabilities.
• LLMs offer potential applications in simulating clients for mental health training.
• Client101 is a novel web platform utilizing LLM-driven chatbots for this purpose.

Purpose of the Study:

• To develop and evaluate a web-based conversational psychotherapy training tool.
• The tool features LLM-driven chatbots simulating clients with mental health conditions.
• To assess the realism and training utility of these AI-powered clients.

Main Methods:

• Developed GPT-4 powered chatbots simulating depression and generalized anxiety disorder.
• Conducted 29 simulated therapy sessions with 16 mental health professionals.
• Administered post-session surveys and analyzed psycholinguistic features using LIWC, comparing to human client data.

Main Results:

• Surveys indicated high participant satisfaction with chatbot realism (93% for anxiety).
• Significant psycholinguistic differences were found between chatbot and human transcripts in 3/8 anxiety and 4/9 depression features.
• Key differentiating features included negations, family, negative emotions, feeling, and health/illness terms.

Conclusions:

• GPT-4 chatbots show promise as effective mental health training tools.
• Participant feedback validates their utility in simulating realistic client interactions.
• Identified psycholinguistic variations provide specific targets for enhancing chatbot realism and training effectiveness.