Medical Large Language Models

Large language models generate text from patterns learned during training and from prompts or retrieved context. In health settings they may draft documentation, summarize records, answer administrative questions, or retrieve guidance. WHO guidance on large multimodal models notes risks including false or inaccurate statements, bias, automation bias, and privacy concerns. A plausible sentence is not proof that it is clinically correct.

The task matters. Summarizing a note for clinician review has different risks from recommending a diagnosis or treatment. Models can omit negation, mix details from records, cite sources that do not support a claim, or fail when prompts are ambiguous. FDA clinical decision-support guidance explains that some software functions fall under device oversight and that users need to independently review the basis for certain recommendations. Teams should determine applicable requirements from intended function rather than assume a general chatbot exemption.

Organizations should test representative cases, measure factual errors and omissions, protect patient data, and provide a human verification step. Use approved environments and least-necessary information. Keep audit logs and incident pathways; do not let generated text silently become the medical record or a treatment order. Clinicians remain accountable for professional decisions. Patients should be told when AI meaningfully contributes to their care. A written use policy should name permitted data, prohibited actions, human sign-off, and a route for reporting errors. Test whether staff can spot unsupported statements before rollout.