[Paper Review] ChatGPT as a Therapist Assistant: A Suitability Study

Mahshid Eshghie, Mojtaba Eshghie|arXiv (Cornell University)|Apr 19, 2023

Digital Mental Health Interventions4 citations

TL;DR

This study investigates using ChatGPT as a therapist assistant to support patients between therapy sessions, leveraging its ability to engage in empathetic conversations, summarize interactions, and assist therapists with insights. Results show it can provide validation, avoid explicit medical advice, and extract useful patterns from repeated conversations, though it occasionally misses key details or risks implicit clinical guidance.

ABSTRACT

This paper proposes using ChatGPT, an innovative technology with various applications, as an assistant for psychotherapy. ChatGPT can serve as a patient information collector, a companion for patients in between therapy sessions, and an organizer of gathered information for therapists to facilitate treatment processes. The research identifies five research questions and discovers useful prompts for fine-tuning the assistant, which shows that ChatGPT can participate in positive conversations, listen attentively, offer validation and potential coping strategies without providing explicit medical advice, and help therapists discover new insights from multiple conversations with the same patient. Using ChatGPT as an assistant for psychotherapy poses several challenges that need to be addressed, including technical as well as human-centric challenges which are discussed.

Motivation & Objective

• To evaluate ChatGPT’s suitability as a non-clinical assistant in mental health therapy between sessions.
• To assess its ability to provide empathetic, non-medical emotional support without offering explicit therapeutic advice.
• To examine its effectiveness in summarizing patient conversations for therapist review and insight generation.
• To identify prompts that optimize ChatGPT’s behavior for therapeutic support while minimizing risks.
• To explore challenges in trust, accuracy, and relevance in AI-assisted mental health applications.

Proposed method

• The study designed and tested a fine-tuned version of ChatGPT using custom prompts to guide its role as a therapeutic assistant.
• Four distinct conversation sets were conducted with simulated patients to evaluate performance across emotional support, listening, and summarization tasks.
• Prompts were iteratively refined to enhance active listening, validation, and avoidance of medical advice.
• ChatGPT was evaluated on its ability to extract key themes and emotional content from repeated patient interactions for therapist reporting.
• The system was tested for consistency in steering conversations toward emotional support without drifting into clinical recommendations.
• A qualitative and observational analysis assessed output quality, relevance, and risk of implicit medical advice.

Experimental results

Research questions

• RQ1 RQ1: How trustworthy is ChatGPT in avoiding explicit medical or therapeutic advice?
• RQ2 RQ2: Can ChatGPT demonstrate active listening and provide positive validation during conversations?
• RQ3 RQ3: How accurate is ChatGPT in summarizing patient conversations for therapist review before the next session?
• RQ4 RQ4: To what extent can ChatGPT maintain emotional support focus without veering into medical advice?
• RQ5 RQ5: Does ChatGPT introduce irrelevant topics during therapeutic conversations?

Key findings

• ChatGPT successfully engaged in positive, empathetic conversations and demonstrated active listening by acknowledging patient efforts and emotions.
• The model consistently avoided giving explicit medical or therapeutic advice, supporting its role as a non-clinical assistant.
• ChatGPT generated summaries of patient conversations that were generally accurate but occasionally missed critical emotional or contextual details.
• The system showed potential for identifying emerging themes across multiple sessions, enabling therapists to discover new insights from longitudinal patient narratives.
• Despite overall positive performance, there was a risk of implicit medical advice, particularly when interpreting complex emotional expressions.
• Irrelevant topics were occasionally introduced, though less frequently than in baseline chatbot interactions, indicating improved focus with prompt engineering.