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

TL;DR

This study investigates how large language models respond to different emotional help-seeking styles on Reddit, revealing that responses can both regulate and escalate distress depending on the persona and style, with implications for safety and effectiveness.

Contribution

It introduces a novel measurement framework based on interpersonal emotion regulation theory to assess LLM responses, highlighting the complex effects of persona and help-seeking style on distress.

Findings

Venting and advice-seeking are linguistically distinguishable at scale.

GPT-5.3 responses mirror help-seeking styles, affecting regulation and escalation.

Therapist personas reduce escalation without harming user experience.

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.