When Patients Say “Do Not Share”: LLMs Misroute Sensitive Information in Clinical Documentation
Year: 2026
Type: preprint
Abstract: Handling and routing sensitive medical information are important aspects of patient care, and patients sometimes ask clinicians not to share particular information. With the growing formal and informal use of AI-generated summaries, we tested whether large language models (LLMs) honor these requests when generating documents for different recipients. We analyzed 20,809 outputs across a seven-model paired evaluation using 120 controlled synthetic cases and 45 patient-authored health queries. Each case had matched versions with and without an explicit request. Despite the request, models disclosed sensitive information in 722 of 2,113 documents generated for unrelated specialists or employers (34.2%). Overall, requests reduced disclosure from 46.6% to 34.2%, a 12.5-percentage-point decrease, leaving sensitive information disclosed in more than one in three such documents. To address this, we evaluated a recipient-conditioned routing workflow designed to omit sensitive information from employer letters and unrelated referrals while preserving it when clinically required. On Llama 3.3 70B, the routing workflow reduced disclosure from 21.7% to 1.2% without increasing harmful omission. Patient requests alone did not provide reliable confidentiality control. Both inappropriate disclosure and excessive omission could undermine trust and safe information sharing. Recipient-specific routing should therefore be evaluated as a core clinical safety function before use.
Source: Research Square
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7204836970