How to review an AI-generated clinical note before you sign it

How should a doctor review an AI-generated clinical note?

Read the note against your own memory of the consultation, not for fluency. Verify identifiers, every figure, all medications and doses, and the plan. Then look specifically for confident statements you did not make. The most dangerous AI errors read perfectly well. Correct, then sign.

Why the signature is the whole point

The Malaysian Medical Council's Guideline 002/2006 on Medical Records and Medical Reports is clear about what a clinical entry has to be. It states that an entry should be "precise and legible", should "record the Date, Time and the Procedure or Treatment, in chronological sequence", and "should be signed by the person making the entry."

Read that last clause again in the context of an AI scribe. The person making the entry is the doctor, not the software. A transcription tool can produce the draft. The moment you sign it, every word in it is your clinical record and your professional responsibility. There is no version of this where the tool carries the liability.

Which is why the review step is not friction to be optimised away. It is the step that makes the whole thing legitimate. Catat drafts your note in roughly two minutes so that you have time to do this properly, not so that you can skip it.

The 60-second review checklist

Run these six in order. Most notes clear in well under a minute once the habit forms.

1. Identity

   Right child, right date, right visit. In a paediatric clinic seeing siblings in one session, this is not a formality.

2. Every number

   Weight, height, temperature, dose, duration, frequency, age in months. Numbers are where speech-to-text fails most often and where the consequences are worst. "Fourteen" and "forty" are one vowel apart.

3. Negatives

   Check that things you ruled out are recorded as ruled out. A dropped "no" turns an excluded symptom into a documented one. Scan specifically for absent negation.

4. Medication and dosing

   Drug name, strength, frequency, and, in paediatrics, whether the dose reconciles with the recorded weight. If it does not, one of the two is wrong.

5. The plan

   Follow-up interval, red flags given to the parent, referrals, investigations ordered. This is the part of the note that gets acted on by someone who was not in the room.

6. The confident stranger

   Anything stated cleanly that you have no memory of saying. Fabricated content does not look uncertain. It reads as the most polished sentence in the note. That is exactly what makes it dangerous.

Where errors actually hide

What makes paediatric consultations harder

General AI scribe advice is written for a two-person consultation. Yours is rarely that.

A paediatric consult is typically three-way: you, the child, and one or two parents, sometimes with a grandparent or helper contributing. The history comes from a proxy who may be anxious, and the patient may be pre-verbal or crying through the examination. Malaysian clinics add another layer: parents commonly switch between Bahasa Malaysia, English and Mandarin or a dialect mid-sentence, and transcription accuracy drops at exactly those switch points.

The practical consequence: pay closest attention to attribution in the Subjective section.

"Mother reports the fever started Tuesday" and "Fever started Tuesday" are different clinical statements.

The first is history from a proxy. The second reads as established fact. Getting this wrong changes how the next clinician reads your note.

Correcting a draft the right way

MMC 002/2006 sets out how corrections work in a paper record: the erroneous statement should be neatly crossed out but remain readable, the correction entered next to it, and signed clearly by the person making the correction. Tampering with clinical notes is to be avoided, because it can be read as covering up a mistake.

The principle carries into digital records: corrections should be visible, attributable and preserved, not silently overwritten. Before you commit to any AI documentation tool, ask the vendor a direct question: does your system keep an audit trail showing what was drafted, what the doctor changed, and when it was signed? If the honest answer is no, you have a governance gap that will surface at the worst possible moment.

Editing the draft before you sign is not a correction to a record. It is authoring. The record begins at your signature.

Before your clinic starts using an AI scribe

• iAgree a house standard for how parents are told recording is happening, and who says it.
• iiDecide who reviews and signs when a locum or trainee sees the patient, before it happens, not after.
• iiiAsk the vendor where the audio and text are processed and stored, how long each is retained, and get it in writing.
• ivSpot-check notes for the first fortnight rather than assuming accuracy, then set a lighter ongoing audit rhythm.
• vHave your data protection obligations reviewed by someone qualified. Malaysia's PDPA framework has been amended recently and provisions have been commencing on staggered dates, so do not rely on a vendor's summary, including ours.