Human Oversight for LLM-Assisted Clinical Documentation

Abstract

Clinical documentation is essential to healthcare delivery but is timeconsuming, cognitively demanding, and prone to error. Large language models (LLMs) can generate draft notes, but may introduce hallucinations, unsupported inferences, or factual distortions that clinicians overlook. This study evaluated whether transparency features improve clinician oversight in LLM-assisted documentation. Using a crossover design, 40 clinicians reviewed AI-generated notes under three conditions: a standard LLM interface, token-level uncertainty highlighting, and uncertainty highlighting combined with direct links to electronic medical record fields. Outcomes included factual errors remaining in finalized notes, editing time, trust, perceived control, and cognitive load. The combined uncertainty-plus-source-link interface performed best, reducing factual errors by 41% and editing time by 18% compared with the standard condition. Uncertainty highlighting alone offered partial benefits but still required manual record verification. These findings indicate that LLM-assisted documentation is safer and more efficient when transparency features support verification, direct attention, and preserve clinician control.