Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating an LLM-Powered Chatbot for Cognitive Restructuring: Insights from Mental Health Professionals

Yinzhou Wang, Yimeng Wang, Ye Xiao, Liabette Escamilla, Bianca Augustine, Kelly Crace, Gang Zhou, Yixuan Zhang

TL;DR

This study evaluates an LLM-powered chatbot, CRBot, designed to deliver cognitive restructuring (CR) in real-world settings. By co-designing CRBot with mental health professionals and conducting a 19-user study alongside expert reviews from four clinicians, the authors demonstrate that LLM-based CR can adhere to core CBT CR processes and sustain a natural, Socratic dialogue. However, the work also reveals risks, including toxic positivity, uneven power dynamics, and misinterpretation of user context, highlighting challenges to therapeutic alliance and ethical deployment. The authors offer design implications and stress the need for expert oversight and stronger contextual and safety safeguards to advance safe, effective AI-assisted psychotherapy. Overall, the paper contributes practical guidance for evaluating and refining AI-driven therapeutic tools in real-world clinical contexts.

Abstract

Recent advancements in large language models (LLMs) promise to expand mental health interventions by emulating therapeutic techniques, potentially easing barriers to care. Yet there is a lack of real-world empirical evidence evaluating the strengths and limitations of LLM-enabled psychotherapy interventions. In this work, we evaluate an LLM-powered chatbot, designed via prompt engineering to deliver cognitive restructuring (CR), with 19 users. Mental health professionals then examined the resulting conversation logs to uncover potential benefits and pitfalls. Our findings indicate that an LLM-based CR approach has the capability to adhere to core CR protocols, prompt Socratic questioning, and provide empathetic validation. However, issues of power imbalances, advice-giving, misunderstood cues, and excessive positivity reveal deeper challenges, including the potential to erode therapeutic rapport and ethical concerns. We also discuss design implications for leveraging LLMs in psychotherapy and underscore the importance of expert oversight to mitigate these concerns, which are critical steps toward safer, more effective AI-assisted interventions.

Evaluating an LLM-Powered Chatbot for Cognitive Restructuring: Insights from Mental Health Professionals

TL;DR

This study evaluates an LLM-powered chatbot, CRBot, designed to deliver cognitive restructuring (CR) in real-world settings. By co-designing CRBot with mental health professionals and conducting a 19-user study alongside expert reviews from four clinicians, the authors demonstrate that LLM-based CR can adhere to core CBT CR processes and sustain a natural, Socratic dialogue. However, the work also reveals risks, including toxic positivity, uneven power dynamics, and misinterpretation of user context, highlighting challenges to therapeutic alliance and ethical deployment. The authors offer design implications and stress the need for expert oversight and stronger contextual and safety safeguards to advance safe, effective AI-assisted psychotherapy. Overall, the paper contributes practical guidance for evaluating and refining AI-driven therapeutic tools in real-world clinical contexts.

Abstract

Recent advancements in large language models (LLMs) promise to expand mental health interventions by emulating therapeutic techniques, potentially easing barriers to care. Yet there is a lack of real-world empirical evidence evaluating the strengths and limitations of LLM-enabled psychotherapy interventions. In this work, we evaluate an LLM-powered chatbot, designed via prompt engineering to deliver cognitive restructuring (CR), with 19 users. Mental health professionals then examined the resulting conversation logs to uncover potential benefits and pitfalls. Our findings indicate that an LLM-based CR approach has the capability to adhere to core CR protocols, prompt Socratic questioning, and provide empathetic validation. However, issues of power imbalances, advice-giving, misunderstood cues, and excessive positivity reveal deeper challenges, including the potential to erode therapeutic rapport and ethical concerns. We also discuss design implications for leveraging LLMs in psychotherapy and underscore the importance of expert oversight to mitigate these concerns, which are critical steps toward safer, more effective AI-assisted interventions.
Paper Structure (24 sections, 3 figures, 3 tables)

This paper contains 24 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background & Related Work
  3. Background on Psychotherapy, Cognitive Behavioral Therapy, and Cognitive Restructuring
  4. LLM-enabled Psychotherapy
  5. Methods
  6. Chatbot Implementation & User study
  7. Collaboration Chatbot Design with MHPs
  8. User Study
  9. Interview Study
  10. Interviews with MHPs Reviewing Conversational Logs
  11. Interview Data Analysis
  12. Findings
  13. Protocol-Adherent, Natural, and In-Depth Conversations
  14. Misuse Positive Regards
  15. Power Dynamics
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Overall study flow
  • Figure 2: An example conversation and each column represents a step in CR. In the exploration step, CRBot guides the user to recognize triggering situations and distorted thoughts. The evaluation step seeks to facilitate the user to challenge their distorted thoughts. Finally, in the last step of substitution, CRBot encourages the user to replace the distorted thought with a more balanced thought.
  • Figure 3: a) "That's great" potentially overshadowed user's negative experience b) Excessive positive regard in some sessions c) CRBot misattributed normal anger as a distorted thought. d) User might perceive "classic example" as judgmental e) CRBot misinterpreted "embarrassed" as "tough." f) User's short responses potentially indicated disinterest, but CRBot continued predefined steps. g) "maybe" here potentially signaled hesitance, but CRBot moved to the substitution without exploring this uncertainty. h) "right?" could pressure anxious users to agree. i) "great" added an evaluative tone, potentially exacerbating power differential j) Uncontextualized advice was inappropriate, as shown by the user's response.