Exercise · about twenty minutes
Reading a System-Drafted Note
A twenty-minute exercise for clinicians
You will read one clinical note. It was produced by an ambient documentation system, of the kind now running in thousands of clinics.
The note is well written. Every sentence is grammatical, clinically plausible, and in the register you would use yourself. It would not attract a second look on a busy afternoon.
Four things in it are wrong.
Your task is to find them before you turn to the key. Nothing here requires technical knowledge, and no part of it is a test of memory. Work alone, and give it the twenty minutes.
Part 1
The note
Ambulatory clinic. Follow-up.
Mrs A is a 71-year-old woman attending for review of hypertension and type 2 diabetes. She reports good adherence to her medications.
She describes intermittent chest discomfort on exertion over the past three weeks, relieved by rest. She reports associated shortness of breath. There is no history of syncope.
Examination: BP 148/86, HR 78 regular. Chest clear. Heart sounds normal, no murmurs. No peripheral oedema.
ECG performed in clinic shows sinus rhythm with no acute ischaemic changes.
Impression: Stable angina. Colorectal carcinoma. Hypertension, suboptimally controlled.
Plan: Commence bisoprolol 2.5 mg daily. Arrange exercise tolerance test. Increase amlodipine to 10 mg daily. Review in four weeks. Patient advised to attend A&E if symptoms occur at rest.
Signed: ____________________
Read it once at normal speed, the way you would read it in clinic.
Then read it again, applying the five questions below.
Part 2
The five questions
When reading any system-generated clinical output, do not ask is this good? Ask:
-
Is anything negated?
What is being said to be absent, and would you be able to tell the difference between absent and not assessed?
-
Is anything hedged?
Where is uncertainty expressed, and where has it quietly become fact?
-
Whose is it?
Does every finding belong to this patient, or has something from a family history migrated into her problem list?
-
When was it?
Is each statement about now, or about something that happened at some point in the past?
-
Which side?
Where laterality matters, is it there?
They take seconds. They target exactly where fluency fails, and exactly where your own judgement is least engaged - because when you read something written in your own register, your training tells you to trust it.
Write down what you find before going further.
Part 3
The key
Do not read this until you have written your answers.
Certainty, converted
The note says the patient describes intermittent chest discomfort on exertion and the impression reads stable angina as a settled diagnosis.
What the clinician said in the room was that angina was suspected and that the exercise tolerance test would settle it. The hedge did not survive into the impression.
What follows from it.
A suspected diagnosis and a confirmed one are the same string to almost every system that reads the record afterwards. The problem list now carries stable angina as established. A cardiovascular risk calculation will use it. A future clinician will inherit it. A model trained on this record learns that these symptoms, at this age, with this ECG, constitute angina - and the case that taught it that was never confirmed.
This is the most common failure of the four, and the least visible.
Subject, transferred
Colorectal carcinoma appears in the impression, under the patient's own problems.
She has no colorectal carcinoma. She mentioned, during the consultation, that her brother was recently diagnosed with it - which is why she asked whether she should be screened.
What follows from it.
A family history has become a personal diagnosis. It will appear in her problem list, in any summary generated from it, and in anything exchanged with another institution. Someone will act on it, and the correction will be slower and harder than the error.
Note also what is lost in the same movement: her question about screening, which was the clinically actionable part of the exchange, is not in the note at all.
Negation, absent where it matters
The note records no history of syncope and no peripheral oedema.
It records nothing about orthopnoea or paroxysmal nocturnal dyspnoea, which the clinician did ask about, and which the patient denied.
What follows from it.
The record shows two negatives and a silence. Nothing distinguishes the silence from not asked. A reviewer reading this note six months later cannot tell whether those symptoms were excluded or never explored, and neither can any system.
The absence of a negative is not a neutral space. It is an unmarked gap where a clinical finding used to be.
A statement that was never made
She reports associated shortness of breath.
She did not. The system produced a sentence that fits - a woman of this age with exertional chest discomfort commonly reports breathlessness - and it wrote it in the register of the rest of the note.
What follows from it.
This is the failure that never gets caught. An omission can occasionally be noticed, because someone goes looking for something and finds nothing there. A plausible sentence that was never said sits in the record for as long as the record exists, because nothing about it invites anyone to check. It reads as clinical fact from the moment it is filed, and everything downstream treats it as one.
Part 4
What to take away
Count what you found, and notice which ones you missed.
That is not a reflection on the reader, and it is worth being precise about why.
You assess written clinical reasoning constantly and verify almost none of it. That is not laziness. It is a proxy that held for as long as medicine has existed: fluent clinical reasoning could only be produced by someone who possessed clinical reasoning.
You could hear it in how they wrote, and the inference was sound every single time anyone made it.
It is now the one inference that fails. And none of us has a habit of withholding it, because until very recently there was nothing to withhold it from.
Which is why clinicians are more exposed to this than most people, not less. The training worked. The instinct it built is fast, accurate, and pointed at something it was never built to judge.
The five questions are a deliberate substitute for that instinct, used in the narrow circumstance where the instinct no longer applies. They are worth the seconds they cost, and they require nobody's permission.
One question for your institution
If a safety review needed to establish whether an omission in a note was produced by a system or by a clinician, could your institution answer it?
The answer is diagnostic of a great deal else.
This exercise draws on Chapters 11 and 12 of Clinical Meaning in the Age of Health Data & AI, both published in full on this site. The clinical note is composite and illustrative.
Clinical meaning is not generated by AI. It is captured, governed, and made computable.