AI scribes are changing clinical training. Note review is becoming a skill of its own.

A new commentary in npj Digital Medicine argues that editing an AI-generated note needs explicit training. Faster documentation and competent review are different outcomes.

The development: a framework for the editing task

A commentary published in npj Digital Medicine on October 12 proposes treating the review of ambient-AI clinical notes as a teachable competency. The authors organize it around detecting errors, reconciling the note with clinical reasoning, retaining independent note-writing ability and calibrating trust in the tool. This is a proposed educational framework, not a new trial showing that a particular scribe improves care. Read the paper by Raka, Van Deusen and Wiberg.

Why review deserves its own attention

The paper identifies a gap between measuring documentation efficiency and measuring whether a reviewer actually catches a note’s errors. That distinction matters when an organization describes its process as clinician-reviewed. A review step can exist in an interface without anyone knowing how well it works.

Consider a synthetic encounter in which a patient describes a previous medication, a current medication and an option discussed for later. A polished draft can read smoothly while placing one of those items in the wrong category. The useful review task is to reconcile the meaning, not simply tidy the prose.

What the paper does not prove

The commentary does not validate a training curriculum through a new controlled deployment. Its framework is a proposal for teaching and evaluation. It also should not be read as evidence that every clinician using ambient documentation is failing to review notes. The contribution is a sharper definition of the task and a way to ask more precise questions about it.

That is a worthwhile development in an industry where a visible draft is easy to demonstrate and review quality is harder to see. Training, supervision and assessment need their own evidence.

What to measure in an AI documentation pilot

Our editorial recommendation is to separate three measures: time to produce a draft, time to finish reviewing it and the quality of the final note. For a training exercise, use invented encounters with known omissions and contradictions. Record which problems reviewers identify and which pass through.

A second useful exercise is a visit completed without the scribe. It helps a team distinguish improvements in the software from changes in the clinician’s own documentation practice. Neither exercise replaces a local clinical validation plan; both make a vague promise about human oversight easier to examine.

The next phase of AI medical documentation will be judged partly on what happens after generation. This month’s paper gives educators and clinical leaders a concrete vocabulary for that work.