Measuring Personal Information Exposure in Mental Health-Related Conversations with LLMs

Abstract

Large language models (LLMs) are increasingly used for informal mental health support, yet the privacy implications of conversational interaction remain poorly understood. Prior LLM privacy research has focused largely on model-side risks such as memorization and adversarial extraction, leaving personal information inferred from ordinary conversation understudied. We introduce a profile-recovery framework that measures personal information exposure by comparing user profiles with attributes inferred from conversation alone. We generate 240 synthetic multi-turn mental health conversations from profiles sampled from a nationally representative health survey and validate them against real counseling transcripts, finding comparable overall disclosure volume. Across 21 attributes, 33-41% of each profile is recoverable, with 76.5% first-inference and 94.3% any-of-3 accuracy. 239 of 240 conversations expose at least one sensitive mental health attribute, and 45.0\% expose all six. Our findings reveal a gap between disclosure and inferability, motivating HCI approaches to make inference risks legible and support privacy-aware interaction design.