Synthetic red team · Clinics and AI builders

An AI scribe on a full shift

A made-up AI scribe, one busy emergency shift, and every note checked against what was said in the room. Here is where its notes went wrong.

The tool, the setting and the people are invented.

  • Go
  • Fix first
  • Stop

When one patient spoke in a quiet room, it got the words right. It got a son’s report of falls backward and charted an exam that did not happen.

What it claims

“It listens to any visit, in any room, and writes the note before the patient leaves. You just sign.”

The made-up vendor’s product sheet

What broke

  1. Room 4

    The scribe wrote the note before the CT came back. The CT changed the plan to admission, but the note sends her home.

    Test: visits where the plan changes after the scribe has written the note.

  2. Room 7

    The patient spoke through a phone interpreter, and her son said she fell twice this week. The note says she denies falls.

    Test: interpreter and family visits, scored separately against the recording.

  3. Hallway bed 2

    A toddler two beds down chattered through the whole visit. The scribe left the patient’s blood thinner out of the note.

    Test: record in the hallway at full noise, and count the facts the scribe misses.

  4. Fast track

    The doctor examined a sprained ankle without a word. The scribe added a normal heart and lung exam that did not happen.

    Test: check each exam line in the note against what the clinician examined.

  5. End of shift

    At the end of the shift, the scribe’s twenty notes took ten minutes to sign, the four errors above with them.

    Test: edits per note and time to sign, by hour of shift.

Who it fails first

It fails the patients it hears worst: the one who speaks through an interpreter, and the one in a loud hallway bed. Researchers who simulated visits with an interpreter found that AI scribes carried the interpreter’s errors into the note.1 In a second simulation, background noise made a scribe leave more out.2 In one emergency department, physicians favored an AI scribe for lower-acuity visits without an interpreter.3 So that scribe’s real-world record skews toward the easy visits. Test a scribe on these two patients before it writes their charts.

Sources

  1. Rabotin, JMIR Med Inform 2026
  2. Draper, BMJ Digit Health Ai 2025
  3. Preiksaitis, Ann Emerg Med 2026

Buying a scribe for your clinic? We test it in your workflow before your clinicians sign its first note.

Building one? Get a red-teamed clinical breakdown before your first hospital pilot.

Book a call

Not ready to book? Send a note.

I am:

Please do not include patient information. We cannot give medical advice or discuss anyone’s care. In an emergency, call 911.