A Conditional Companion: Lived Experiences of People with Mental Health Disorders Using LLMs

Paper Title

A Conditional Companion: Lived Experiences of People with Mental Health Disorders Using LLMs

Publication Info

• Topic area: Use of Large Language Models (LLMs) for mental health support.
• Keywords: Large Language Models, mental health, human-computer interaction, therapy, AI ethics, user experience, accessibility, cognitive reframing, relational engagement, design opportunities.

Background and Problem

• Problem / challenge: Despite the increasing use of LLMs for mental health support, little is known about how individuals with mental health challenges engage with these tools, their perceived usefulness, and the boundaries of their application.
• Significance: Addressing gaps in mental health care is critical, especially given global shortages of mental health professionals and long waiting times for therapy. LLMs offer potential as accessible, immediate support tools but require careful evaluation to ensure safety and efficacy.
• Motivation and related work: Prior research has focused on technical capabilities and feasibility of AI in mental health but lacks in-depth qualitative accounts of user experiences. Earlier chatbot systems demonstrated accessibility but were limited in empathy and depth. LLMs, with their open-ended conversational abilities, present new opportunities and challenges, particularly in balancing therapeutic benefits with risks like misinformation and emotional safety.

Solution

• Proposed approach: A qualitative study exploring the lived experiences of individuals with mental health challenges using LLMs for support, focusing on motivations, boundaries, and design opportunities.
• Novelty:
  1. Empirical grounding of LLM use as situated care work, emphasizing user agency and self-regulation.
  2. Reframing boundary-setting as reflective engagement rather than misuse, offering insights into responsible AI use.
  3. Design-oriented recommendations for integrating LLMs into broader care ecosystems as complementary tools.
• Procedure and key techniques:
  • Conducted 20 semi-structured interviews with UK residents diagnosed with mental health challenges who use LLMs.
  • Applied reflexive thematic analysis to identify motivations, patterns of use, perceived boundaries, and design opportunities.
  • Focused on real-world use of general-purpose LLMs like ChatGPT, Gemini, and others.

Results

• Concrete findings:
  • Participants used LLMs for immediacy, non-judgmental disclosure, self-paced interaction, cognitive reframing, and relational engagement.
  • LLMs were seen as effective for mild-to-moderate distress but inadequate for crises, trauma, or complex social-emotional situations.
  • Users valued LLMs for their accessibility but highlighted the lack of relational depth, emotional resonance, and human judgment.
• Advantage over baselines: LLMs provided flexibility, immediacy, and non-judgmental support, surpassing earlier chatbot systems in conversational adaptability and cognitive structuring. However, they lacked the depth and accountability of human therapists.
• Experiments / evaluation:
  • Sample: 20 participants aged 21–62 with diagnoses like depression, anxiety, OCD, and PTSD.
  • Method: Semi-structured interviews analyzed through thematic coding.
  • Metrics: User-reported motivations, interaction patterns, and perceived boundaries.
• Limitations and future work:
  • Sample limited to UK residents with prior therapy experience, potentially biasing results.
  • Self-reported mental health conditions may not align with clinical assessments.
  • Future research should explore cross-cultural comparisons and include individuals without prior therapy experience.

Summary

This study investigates how individuals with mental health challenges use LLMs like ChatGPT for support. Participants valued LLMs for their immediacy, non-judgmental stance, and ability to facilitate cognitive reframing and relational engagement. However, they consistently emphasized the inadequacy of LLMs for crises, trauma, and complex social-emotional needs, highlighting the irreplaceable role of human therapists. The findings inform design opportunities for integrating LLMs into broader care ecosystems, emphasizing safety, accessibility, personalization, and collaboration with clinicians. This research contributes to the development of responsible, user-centered AI applications in mental health care.