Study: Clinicians add more hedging to AI-drafted clinical notes

A new preprint study by Yiliang Zhou and colleagues examines how clinicians modify hedging language when editing ambient AI-generated clinical notes. The researchers analyzed 62,811 paired note sections from multiple AI vendors across various clinical specialties, comparing pre-edit AI drafts to post-edit final notes. They found that hedging terms were more frequently introduced into previously non-hedged text than removed from previously hedged text, and the overall post-edit text contained more hedging mentions than the pre-edit drafts. Directionality analysis revealed a significant tendency toward greater uncertainty in the replacement edits, suggesting clinicians systematically push the language of AI drafts toward less certainty.

The study further broke down results by vendor and specialty, showing substantial heterogeneity in hedging prevalence, change in hedging mentions, and directionality of edits. This suggests that the behavior of clinicians modifying AI-drafted notes is not uniform—it depends on both the specific AI system used and the medical context. The findings have important implications for ambient AI documentation tools (like DAX Copilot or Abridge), which are increasingly deployed to reduce physician burnout. If clinicians consistently add hedging language, the final notes may actually be less decisive than AI drafts, potentially impacting clinical decision-making, billing, and legal documentation. The study is available on arXiv (2606.00018).