Understanding clinician edits to ambient AI draft notes: a feasibility analysis using large language models.

Reviewing clinician edits to ambient AI-drafted notes can help assess AI reliability. This pre-print study tests if LLMs can learn from clinician revisions. Medication and symptom models performed well at categizing clinician edits, but diagnosis, orders/tests/procedures, and social history edits were less precise. Prompt engineering worked helped categorize straightforward edits, but complex cases need triage, with human review of labeled edits.