What a paediatric AI scribe gets wrong, and why the parent's account is the hard part

An AI scribe is very good at turning speech into a tidy note. It is not listening clinically, and it does not know what mattered in the room. That gap is where its mistakes live, and the mistakes are more predictable than they first appear.

The errors cluster in five places

Across a clinic day, the same categories come up again and again:

• Misheard numbers. Doses, weights, temperatures and intervals. This is where speech-to-text fails most often, and where the consequences are worst.
• Dropped negatives. A missing "no" quietly turns a symptom you excluded into one you documented.
• Speaker confusion. In a room with a parent, a child and you, the note can credit the wrong person with an account.
• Fabricated detail. A sentence that reads cleanly and sounds clinical, but that you never actually said.
• Lost code-switching. Accuracy dips at the exact point a sentence moves between Bahasa Malaysia, English and Mandarin.

None of these look like errors on the page. That is the whole problem. The most dangerous line in a draft is usually the most fluent one.

Why paediatrics is the hard case

Most AI scribe advice assumes a two-person consultation: doctor and patient. Yours is rarely that.

"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. A paediatric history comes through a parent who may be anxious, about a patient who may be pre-verbal or crying through the examination. The account is second-hand by design, and an AI scribe cannot tell the difference between what was reported and what was observed unless the transcript makes it obvious.

So the section that needs your closest attention is the Subjective. Check who is credited with each statement before you sign.

What this means in practice

It does not mean the tool is not worth using. It means the review is not optional. A draft that arrives in two minutes is only a saving if the minute you spend checking it is a real one.

If you want the full method, we wrote it up as a checklist: how to review an AI-generated clinical note before you sign it. The short version is that you read the note against your own memory of the consultation, not for how well it reads.

Catat drafts the note and hands it back. The clinical judgement, and the signature, stay yours.

← All posts