An AI scribe dropped one word and reversed a diagnosis

A patient in England was told they had demyelination, the nerve damage behind conditions including multiple sclerosis. The test result had said null demyelination. An AI scribe dropped the word that reversed the meaning, and the patient caught it.

Healthwatch England, the statutory patient champion for the NHS, published that case on 31 August alongside others. Twenty-seven different scribe products are in use across the service.

“I can state with certainty that I find meaningful errors far more often in AI consultations than in those typed out by colleagues.”

— Dr Mary Gibbs, out-of-hours GP, The Guardian, 7 September 2026

Quote source: Dr Mary Gibbs, out-of-hours GP in Manchester, The Guardian letters, 7 September 2026

The time saving goes to the second reader

Out-of-hours GP Mary Gibbs wrote to The Guardian about the part of the pitch that gets least scrutiny. Her patients are often assessed by someone else on the phone first, and she reads that triage note before calling them in.

When the note came from a scribe, she writes, the consultation is invariably long, with duplications and sometimes contradictions, and the history she takes herself often differs from what is written down. So the record gets checked twice:

1. The first clinician has to proofread a long, repetitive account they did not write.
2. The second clinician cannot trust it and takes the history again.
3. The record is read three times and believed none of them.

Gibbs makes the point that blaming the first clinician misses how proofreading works. Anyone who has tried to check their own text under time pressure knows the errors survive. A tool that produces a draft nobody can trust has not removed the work, it has moved the work to the person reading.

The errors that are not funny

Healthwatch found scribes swapping a prescribed medicine for a similarly named one, and clinicians reported software inventing medications nobody mentioned. In the cases that surfaced, the patient noticed before the doctor did.

Debbie Cameron, writing in the same letters column, recalls a consultant who trialled voice recognition years ago and abandoned it when a recommendation of two weeks on lansoprazole came out as a two-week holiday on Lanzarote. That failure was funny because it was obvious. A dropped null is not.

Nobody calls them medical devices

The Medicines and Healthcare products Regulatory Agency has not classified AI scribes as medical devices. Twenty-seven products are transcribing live consultations across the NHS without the testing regime that would apply if they were labelled as what they functionally are.

Gibbs closes on the part that outlasts any individual error. Turning a patient's rambling account into something a colleague can absorb in thirty seconds is a skill that takes years, and clinicians who stop practising it will lose it.

We wrote yesterday about a trade that AI removed outright, and on Monday about industry bodies reassuring designers. Medicine sits in a third position: the tool did not take the job, and it changed what the job consists of, from writing a summary to auditing one.

This piece is informational, not a recommendation to buy, sell, or hold any asset.