Interpreting AI-Mediated Support: Understanding Its Effectiveness in Social Media–Induced Anxiety

Abstract

Algorithmically curated social media feeds can provoke perceived anxiety in everyday use, yet little is known about how AI can provide effective support in these contexts. We present TriggerDetector, an LLM-powered system that infers plausible anxiety triggers from user-encountered posts and offers coping suggestions. Following a validation study, we evaluate the system through an in-the-wild deployment with 241 users engaging with their own anxiety-inducing posts. Results show that many users reported reduced anxiety after use (p <.001, Cohen’s d = 0.87), though outcomes varied substantially across interaction instances (each corresponding to one uploaded post), including instances where support was ineffective or increased anxiety. Further analysis indicates that effectiveness depends less on trigger identification and more on how AI-generated suggestions are interpreted and enacted. Interviews (n = 10) reveal how value misalignment, limited actionability, and problematic interpretations shape these outcomes. We contribute empirical insights into when and how AI-mediated support is effective, and derive design implications for safer and more context-sensitive support.