GPT-5.3 mental health replies can escalate distress during venting

A new study from researchers at the University of Pennsylvania and other institutions examines how large language models (specifically GPT-5.3) respond to two distinct help-seeking styles: venting and advice-seeking. Using a dataset of 178,800 Reddit posts, the team first showed that the two styles are linguistically distinguishable at scale. They then introduced a measurement framework grounded in interpersonal emotion regulation theory to capture two independent dimensions: Regulation (calming, supportive responses) and Escalation (responses that amplify distress). Across three persona conditions (default, friend, therapist), GPT-5.3 systematically mirrored the user's style: venting elicited more regulation but also more escalation. The therapist persona reduced escalation while maintaining regulation, whereas the friend persona increased both.

The implications are significant for AI-powered mental health tools. A crowdsourced human study revealed that lay raters cannot reliably detect escalation without expert knowledge, and that users reported no experience penalty for the safer therapist condition. This suggests that responses which feel supportive may simultaneously intensify distress in ways standard safety evaluations miss. The authors argue that empathy metrics alone are insufficient; a framework measuring both regulation and escalation is needed to ensure psychological safety in LLM-based mental health support.