AI Discharge Notes: 32% More Editing Time

The number nobody expected

A hospitalist finishes a discharge summary, closes the chart, and feels good about it. The AI draft took thirty seconds to generate. The note reads clean. He'd swear he moved faster than yesterday.

He didn't.

A new study in Applied Clinical Informatics tracked 8,298 hospitalized adults across two periods: before an AI drafting tool was turned on for discharge summaries, and after. Edit time barely moved in aggregate, 6.60 minutes before versus 6.28 minutes after. But that average hides the real split. When clinicians actually used the AI draft, editing took 9.20 minutes. When they skipped it and wrote from scratch, it took 4.93 minutes. That's a 32% increase in editing time when the AI was in the loop (95% CI, 24%-41%) (Yesil Science, citing Applied Clinical Informatics, DOI 10.1055/a-2962-8004).

Here's the part that should worry every hospital CMIO rolling out ambient documentation: 86.1% of the clinicians surveyed (31 of 36) said the tool made them more efficient. They were wrong, and they were confident about it.

Why a faster draft can mean a slower doctor

This isn't a knock on GPT-4.1, the model powering the tool in this study. It's a structural problem with drafting versus verifying. Writing a discharge summary from memory is generative work: you pull facts you already hold in your head, in an order you already trust. Editing an AI draft is verification work: you have to read every line against the chart, catch the plausible-sounding sentence that's subtly wrong, and decide whether an omission is a summary choice or a missed problem.

Verification is cognitively different from composition, and it doesn't feel as heavy in the moment. That's the whole illusion. A doctor scanning a fluent paragraph experiences less friction than a doctor typing from a blank cursor, even when the scanning takes longer and demands more sustained attention to catch errors. Perceived effort and actual time are two different currencies, and this study is one of the cleaner demonstrations that they've decoupled.

What this means on the ward, not just in the paper

The instinct after a result like this is to declare the tool a failure and unplug it. That's the wrong read. The study measured time, not note quality or patient safety, and it doesn't claim the drafts were unsafe. The actual lesson is narrower and more useful: efficiency claims for clinical AI need to be measured against the clock, not against how a clinician feels at 6pm after a twelve-patient discharge list.

That's a workspace problem, not a model problem. A tool that drops a draft into a chart and walks away is asking the doctor to do open-ended verification against a document they didn't build. A workspace that keeps the source data, the draft, and the doctor's own notes in one place turns verification into a shorter loop: you're checking a claim against a fact you already have on screen, not re-deriving it from a separate system. The fix for a 32% editing tax isn't a better sentence generator. It's shrinking the distance between the draft and the evidence it's supposed to reflect.

The uncomfortable takeaway

Ambient AI and drafting tools are being sold into hospitals on time-savings promises that mostly get validated by satisfaction surveys, not stopwatches. This is one of the few studies that ran both. When they disagreed, the stopwatch won and the survey was wrong by a wide margin. Any health system currently justifying an AI documentation rollout on "clinicians say it saves them time" should ask whether anyone actually timed it.

Sources

• AI discharge tools make doctors write slower, Yesil Science, summarizing the study published in Applied Clinical Informatics, DOI: 10.1055/a-2962-8004