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

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

1. Problem Statement

There is a growing global demand for mental health support, exacerbated by systemic barriers such as professional shortages, high costs, and long waiting lists (e.g., months-long delays in the UK's NHS). While Large Language Models (LLMs) like ChatGPT and Claude are being proposed as scalable solutions due to their 24/7 availability and conversational adaptability, a significant research gap exists. Most current literature focuses on the technical capabilities (fine-tuning, diagnostic accuracy, and risk detection) rather than the lived experiences of the users. There is a lack of understanding regarding why individuals turn to general-purpose LLMs, how they navigate the risks of these "imperfect solutions," and where they perceive the boundaries between AI-mediated support and professional human care.

2. Methodology

The researchers employed a qualitative, human-centered approach to capture nuanced user perceptions:

• Study Design: A qualitative study consisting of 20 semi-structured, retrospective interviews.
• Participants: 20 individuals residing in the UK, aged 21–62, who self-reported diagnoses of mental health conditions (e.g., Anxiety, Depression, OCD, PTSD, ADHD) and had actively used LLMs for mental health support.
• Exclusion Criteria: To ensure data reliability and safety, participants with severe conditions involving potential loss of contact with reality (e.g., schizophrenia, bipolar disorder, or psychotic disorders) were excluded.
• Data Analysis: The researchers used reflexive thematic analysis, an iterative six-phase process involving coding, reviewing, and consolidating themes through weekly analytic meetings to reach consensus.

3. Key Contributions

The paper provides three primary contributions to the Human-Computer Interaction (HCI) and AI communities:

• Empirical Grounding of "Situated Care Work": It reframes LLM use not as passive adoption, but as an active, ongoing process of sense-making, risk assessment, and self-regulation by the user.
• Conceptualization of "Boundary-Setting": It introduces an analytic account of how users actively establish and revise boundaries around AI use based on therapeutic experience, reframing "limitations" as a form of competent, reflexive engagement.
• Design and Governance Frameworks: It provides actionable insights for positioning LLMs as complementary tools within existing care ecosystems rather than replacements for human professionals.

4. Results

The findings are categorized into three research questions:

RQ1: Motivations and Patterns of Use
Users engage with LLMs through five primary "needs":

• Immediacy: Accessing instant support during "off-hours" or while waiting for professional appointments.
• Non-judgmental Disclosure: Using the AI as a safe space to vent stigmatized or "embarrassing" thoughts without fear of social judgment.
• Self-Paced Disclosure: Controlling the tempo and depth of conversation, which contrasts with the time-constrained nature of traditional therapy.
• Sense-making of Inner Thoughts: Utilizing LLMs for cognitive restructuring (similar to CBT) to break down anxiety spirals into manageable steps.
• Relational Engagement: Anthropomorphizing the AI to fulfill a need for companionship and "friend-like" presence during loneliness.

RQ2: Boundaries of Support
Participants identified a "crisis/complexity threshold" where LLMs become inadequate:

• Crisis/Emergency: LLMs are deemed unsafe for suicidal ideation or acute trauma.
• Social/Emotional Complexity: LLMs struggle with the nuances of interpersonal relationships and socially embedded dilemmas.
• Relational Depth: LLMs lack the emotional resonance, continuity, and ability to detect subtle non-verbal cues that a human therapist provides.

RQ3: Design Opportunities
Users envisioned several functional improvements:

• Safety: Clearer escalation pathways and crisis signposting.
• Accessibility: Shorter, multimodal (visual/audio) outputs to reduce cognitive load during distress.
• Personalization: User-controlled memory (selective saving/deleting of history) and customizable interaction tones.
• Integration: Connecting LLMs with wearables (e.g., Fitbit) and clinical notes to create a holistic view of wellbeing.

5. Significance

This research challenges the "techno-optimist" view that AI can replace therapists and the "paternalistic" view that users are naive. Instead, it reveals a pragmatic acceptance of imperfect solutions: users recognize the flaws and risks of LLMs but utilize them as a rational response to systemic failures in healthcare.

The study carries significant implications for AI Governance, arguing that the responsibility for safety should not rest solely on the individual user (who is already in a vulnerable state) but must shift toward corporate accountability, rigorous clinical benchmarks, and regulatory oversight (e.g., treating LLMs as medical devices). It moves the conversation from "Can AI do therapy?" to "How can AI safely support the existing mental health ecosystem?"