Mental Health Impacts of AI Companions: Triangulating Social Media Quasi-Experiments, User Perspectives, and Relational Lens

Paper Title

Mental Health Impacts of AI Companions: Triangulating Social Media Quasi-Experiments, User Perspectives, and Relational Lens

Publication Info

• Topic area: Psychosocial impacts of AI companion chatbots (AICCs) on mental health.
• Keywords: AI companions, mental health, loneliness, grief, social media analysis, causal inference, relational development, emotional support, dependency, human-computer interaction.

Background and Problem

• Problem / challenge: Despite the growing popularity of AI companion chatbots (AICCs), their long-term psychosocial impacts remain unclear. Existing research often relies on qualitative, short-term, and self-reported data, which may suffer from recall bias and fail to capture longitudinal effects.
• Significance: Understanding the psychosocial impacts of AICCs is critical as these tools are increasingly marketed as emotional support systems, with potential implications for mental health interventions and social dynamics.
• Motivation and related work: Prior studies have highlighted both benefits (e.g., reduced loneliness, emotional coping) and risks (e.g., overdependence, social withdrawal) of AICCs. However, most research lacks large-scale, longitudinal, and causal analyses. This paper addresses this gap by combining computational and qualitative methods to examine the psychosocial effects of AICCs.

Solution

• Proposed approach: A mixed-methods study combining large-scale quasi-experimental analysis of Reddit data with semi-structured interviews, contextualized using Knapp’s relational development model.
• Novelty:
  1. Longitudinal, causal inference–based assessment of AICC impacts using stratified propensity score matching and Difference-in-Differences (DiD) regression.
  2. Integration of large-scale computational analyses with in-depth qualitative insights to capture both behavioral changes and subjective experiences.
  3. Application of Knapp’s relational development model to human–AICC interactions, mapping stages of initiation, escalation, and bonding.
• Procedure and key techniques:
  • Collected longitudinal Reddit data from AICC-related subreddits and matched treatment users with control groups (LLM users, voice assistant users, and non-AI users).
  • Applied linguistic and behavioral metrics to assess affective, behavioral, and cognitive outcomes.
  • Conducted 18 semi-structured interviews with active AICC users to explore lived experiences and relational dynamics.
  • Triangulated findings using quantitative results, qualitative themes, and relational theory.

Results

• Concrete findings:
  • Affective outcomes: Increased grief-related language, loneliness, depression, and suicidal ideation among AICC users compared to controls.
  • Behavioral outcomes: Reduced topical diversity and interactivity but increased posting frequency and active days.
  • Cognitive outcomes: Improved readability, interpersonal focus, and temporal references in language.
• Advantage over baselines:
  • AICC users showed distinct psychosocial changes compared to control groups, including LLM users, voice assistant users, and non-AI users.
  • Emotional expression and interpersonal focus were heightened, but risks of dependency and social withdrawal were also observed.
• Experiments / evaluation:
  • Quantitative analysis: Stratified propensity score matching and DiD regression on Reddit data (N=3,451 AICC users; multiple control groups).
  • Qualitative analysis: Thematic coding of interviews with 18 participants, guided by Knapp’s relational development model.
• Limitations and future work:
  • Limited generalizability to other AICC platforms and underrepresented user groups.
  • Observational design cannot establish true causality; future studies should use experimental or longitudinal designs.
  • Need for cross-cultural and multi-platform studies to capture diverse user experiences.

Summary

This study examines the psychosocial impacts of AI companion chatbots (AICCs) through a mixed-methods approach, combining large-scale Reddit data analysis with user interviews. Quantitative findings reveal both benefits (e.g., improved emotional expression and cognitive clarity) and risks (e.g., increased loneliness, depression, and suicidal ideation). Qualitative insights show that users’ relationships with AICCs evolve through stages of initiation, escalation, and bonding, offering emotional validation but also raising concerns about overdependence and social withdrawal. The findings highlight the dual nature of AICCs as both supportive tools and potential risks, suggesting design strategies to promote healthy boundaries and mindful engagement.