AI notetakers go wrong in three predictable places

Something quietly joined most people’s meetings this year. It listens, it writes the notes, and by the time you have got your coat on it has emailed everyone a summary. Almost nobody has asked it the obvious question: what did you miss?

Two pieces of UK research published this year happen to answer that, and they come from a setting where the stakes are unusually clear — a GP’s consulting room. The findings generalise better than you would expect to a sales call, a site visit or a disciplinary meeting.

The short version

• Errors are not random. They cluster. Among UK GPs using these tools, errors were most frequent in multiparty consultations (38%), complex histories (35%) and non-English encounters (31%).
• Roughly one user in seven reported errors with significant-to-critical implications — 14% of the 141 users in the survey, so 20 doctors.
• The adoption figure being repeated this week is 40%. The biggest survey we can read says 14% are currently using them, with a further 39% intending to.
• What the tools miss is everything that was not said out loud: facial expressions, gestures, emotional state.
• Being recorded changes what people are willing to tell you.
• And the benefits are real: 80% reported less time on documentation. This is not an argument for switching them off.

Where they go wrong, and it is predictable

The useful research here is a nationwide survey of UK general practitioners by Charlotte Blease and colleagues, published in BMJ Health & Care Informatics, with fieldwork in August 2025 and 1,003 respondents recruited through Doctors.net.uk.

Of those 1,003, 141 were currently using an ambient AI scribe. Every figure that follows is out of those 141 users, not out of all GPs — which matters, because it is the difference between “a third of doctors” and “a third of the doctors who actually use one”.

Among those users:

• 32% (45 of 141) reported errors often or always.
• 14% (20 of 141) reported errors with significant-to-critical implications.
• Errors were “most frequent in multiparty consultations (38%), complex histories (35%) and non-English encounters (31%)”.

That last line is the one to write down. The failure modes are not mysterious and they are not evenly spread. More than one person talking. A story that does not run in a straight line. A language the tool was not really built around.

Every one of those has an exact equivalent in ordinary business meetings: the three-way call where people talk over each other, the client whose situation takes twenty minutes to explain properly, and the supplier conversation in accented or second-language English. We have written before about how much accent affects machine transcription, and this is the same weakness showing up in a different tool.

The number under the number

Here is where it gets awkward, and we would rather show you the working than quietly pick a side.

The University of Edinburgh’s news release, published on 4 September 2026, says that “Some 40 per cent of GPs in the UK have said they used ambient scribes”, in a sentence that goes on to explain what the tools are. That figure is being repeated this week as though four in ten GPs are using one now.

The largest UK survey we have been able to read says something different. In August 2025, of 1,003 GPs:

• 14% (141) were currently using ambient AI scribes.
• 39% (396) intended to adopt them soon.
• 46% (466) had no plans to use them.

Those three add up to 1,003, so nothing is missing. But 14% is not 40%, and 14% plus 39% is 53%, which is not 40% either.

We are not saying Edinburgh is wrong. Their sentence says GPs “have said they used” the tools, which is a broader question than current use, and they may be citing a source we have not found. Their review is published in BMJ Digital Health and AI, and the journal’s site returns a challenge page we cannot get past, so we have not read it. What we can tell you is that the readable evidence puts current adoption a good deal lower than the headline, and that an intention to adopt is not an adoption. If you are making a decision on the basis of “everyone is doing this already”, that is worth knowing.

What a recording does not contain

The Edinburgh review — a review of 27 articles in the scientific and medical literature, rather than new fieldwork, which is worth being clear about — is about the subtler cost. Their release puts it plainly: the tools can miss “important aspects of consultations such as facial expressions, gestures and emotional state”, and AI summaries “tend to prioritise clinical information, which can come at the expense of the patients’ stories and experiences of their illness”.

Two further findings deserve more attention than they will get.

First, recording changes the conversation. Patients “may be hesitant to reveal sensitive information, such as substance abuse, domestic abuse or mental health struggles, when they know the consultation is being recorded and processed by AI”. Whatever your business is, there is a version of this: the reason a deadline slipped, the real objection behind a polite one, the thing an employee will say to you but not to a transcript.

Second, you may stop remembering. The review describes clinicians who “do not recognise their own notes or remember the patient the next time they see them”. Writing something down is not only a record; it is part of how you think about it. Hand that over and you get the time back, but you lose some of the thinking.

Dr Lucas Seuren of Edinburgh’s Centre for Biomedicine, Self and Society, quoted in the release, puts the trade-off well: “Many clinicians are excited about ambient AI scribes, because they promise to cut down on paperwork. But the experiences of patients are poorly considered, and there are real risks that the patients’ stories are lost.”

Why this is really a lesson about machines and words

Strip out the healthcare and a general rule is left standing: an AI system only gets what was put into words. Everything else in the room — the pause, the glance, the thing someone pointed at, the fact that nobody objected — simply is not in the data.

That is the same rule that governs how machines see your business from the outside. An answer engine describing you to a customer has your words and nothing else. The atmosphere of your shop, the care in your work, the thing everyone locally knows about you: none of it exists unless somebody wrote it in a sentence. It is exactly the point we made about writing descriptions that work for people and machines at once — if a fact only lives in a photograph, the machine does not have it.

The notetaker in your meetings and the AI describing you to a stranger are running the same limitation. One of them you can watch operating in real time, which makes it a good place to learn what the other one is doing.

What to do on Monday

1. Check the notes from meetings that match the three risky shapes — several people talking, a complicated history, or a conversation not held in English. Not every meeting. Those ones.
2. Say out loud the things you want recorded. If the decision was made by everyone nodding, nobody made it as far as the transcript is concerned.
3. Tell people it is running. In the survey, 63% of users routinely sought consent and no more than one patient in ten declined. Asking costs you very little and it is the honest thing to do.
4. Read your own notes once before you rely on them. Especially the first few weeks, while you still remember what actually happened.
5. Decide which conversations it does not attend. There will be some.

What argues the other way

The tools are working, on the users’ own account. 80% reported reduced time on documentation, 70% reported reduced cognitive load, and 55% rated the output better than their standard notes. A post that only listed the failures would be a dishonest reading of the same survey.

This is doctors, not you. A consultation is a uniquely demanding thing to summarise: emotional, high-stakes, and full of information the patient is reluctant to volunteer. Your Tuesday project call is not that. The error rates almost certainly do not transfer — but the shape of the errors is about how the technology works, not about medicine, and that is the part we have leant on.

And the fieldwork is a year old. The survey ran in August 2025 and these tools change quickly, so both the adoption figure and the error rates may be out of date in either direction. That is an argument for re-checking rather than for ignoring it, and it is also, incidentally, why the 40% and the 14% may both be perfectly true of different moments.

Sources, and what we checked

• The survey figures — Blease C, Kharko A, Sanchez CG, Navarro D, McMillan B, Gaab J, Locher C, Coiera E, “Ambient AI in primary care: an exploratory mixed methods survey of UK general practitioners”, BMJ Health & Care Informatics, first published 1 July 2026 (DOI 10.1136/bmjhci-2025-101847). Read from the published abstract via Europe PMC. Fieldwork August 2025; 1,003 respondents recruited via Doctors.net.uk; 141 current users. All percentages above are recomputed from the counts given, and the three adoption categories sum to 1,003. The abstract was re-queried on 6 September before publishing and came back identical, with all fourteen figures unchanged.
• The review findings and the 40% figure — University of Edinburgh news release, “AI scribes may fail to capture patients’ experiences”, published Friday 4 September 2026, read in full. It describes a review of 27 articles, published in BMJ Digital Health and AI. Re-fetched on 6 September before publishing: unchanged to the character, including the 40% sentence, so the gap described above is not something that has since been corrected.
• What we could not read — the review paper itself. BMJ Digital Health and AI returned an HTTP 403 challenge page to us on 4 September 2026 and again on twice-repeated attempts on 6 September, so everything attributed to the review comes from the authors’ own university release rather than from the paper. We have not repeated any figure or phrase we could not find in a source we actually read.

Every quotation above was checked word for word against the source we read it in. Percentages described as ours are recomputed from the published counts; where the two sources disagree, both numbers are printed.