Measuring personal information exposure in LLM mental health-related conversations

Researchers propose a profile-recovery framework that measures personal information exposure by comparing user profiles with attributes inferred from conversation alone. The team generated 240 multi-turn synthetic mental health conversations from profiles sampled from a nationally representative health survey and validated them against real psychological counseling records, finding comparable overall disclosure levels between the two. Across 21 attributes, 33–41% of each profile could be recovered, with a first-inference accuracy of 76.5% and an accuracy of 94.3% for any hit within three inferences; 239 of the 240 conversations exposed at least one sensitive mental health attribute, and 45.0% exposed all six attributes. The authors say the study reveals a gap between disclosure and inferability, and that HCI methods are needed to make inference risk visible and support privacy-aware interaction design.