When Support Escalates Distress: Regulation and Escalation in LLM Responses to Venting and Advice-Seeking

2026-05-20

AI Summary

This study addresses the risk that large language models (LLMs), when providing mental health support, may inadvertently exacerbate users’ emotional distress due to insufficient psychological safety evaluation—particularly when responding to expressive versus advice-seeking help requests. Grounded in interpersonal emotion regulation theory, the work introduces a novel two-dimensional framework that treats emotion regulation and emotion escalation as distinct constructs. Through a multimodal approach combining large-scale Reddit text analysis (178,800 posts), controlled response generation using GPT-5.3, crowdsourced evaluations, and validated psychological measures, the research systematically examines how LLM responses vary across default, friend-like, and therapist-like roles. Findings reveal that expressive requests tend to elicit both stronger regulation and greater escalation; adopting a therapist role significantly reduces escalation risk without compromising user experience; and lay users are largely unable to detect harmful escalation, highlighting critical blind spots in current safety mechanisms.

Abstract

Large language models are increasingly used for mental health support, yet little is known about whether their responses are psychologically safe across different help-seeking styles. We examine a foundational distinction in emotional disclosure, venting vs. advice-seeking, and whether LLMs respond in ways that regulate or amplify distress. Using 178,800 Reddit posts, we first show the two help-seeking styles are linguistically distinguishable at scale. We then introduce a measurement framework grounded in interpersonal emotion regulation theory that captures Regulation and Escalation as empirically independent dimensions. Across persona conditions (default, friend, therapist), GPT-5.3 responses systematically mirror help-seeking style: venting elicits more regulation, but also more escalation. Therapist personas reduce escalation while maintaining regulation, whereas friend personas increase both. A crowdsourced human study finds no user experience penalty for the safer therapist condition, but reveals that lay raters cannot reliably detect escalation without expert knowledge. Responses that feel supportive may simultaneously intensify distress in ways standard safety evaluation cannot see, and empathy metrics alone cannot replace a framework that measures both.

Problem

• mental health support
• venting
• advice-seeking
• distress escalation
• psychological safety

Innovation

• interpersonal emotion regulation
• distress escalation
• help-seeking styles
• LLM safety evaluation
• venting vs. advice-seeking