Image by https://www.magnific.com/

Yesterday, The Guardian reported on a Healthwatch England warning about errors made by AI scribes - tools that listen to our medical consultations and generate notes for our doctors in real-time. In the survey, patients reported wrong medication names, incorrect diagnoses and omissions from their records. I was quoted in the piece saying that “AI can and does make mistakes.”

The Healthwatch survey is very important because we have very few surveys of patients’ opinions and experiences about AI, including doctors using these tools. Therefore the repotr is definitely to be welcomed.

But it is important to contextualise this finding.

AI can and does make mistakes. In our own research in 2025 of GPs who used ambient AI, doctors also believed errors were more common in consultations with more than one person, those with complex patient histories, and among patients whose first language isn’t English. Such errors could drive health inequities.

But it is also critical to ask: compared to what?

The fact is, doctors can and do make mistakes without AI. And it is certainly possible the error rate is worse with ambient AI scribes. But more than half the doctors in our 2025 survey believed that their ambient AI records were better than standard records.

To err is human, and we know that with standard approaches doctors do make some of the same errors: wrong diagnosis, wrong side of body documented, even the wrong patient’s notes. Yet there has been much less attention to this problem than before AI was introduced. A large pre-AI study led by my colleague Sigall Bell found around one in five patients who read their notes reported finding an error; among those, around 40% believed the error was serious.

Why do doctors make mistakes? Writing up documentation after the visit - often at the weekend - means errors can and do slip in. Doctors are notorious for their bad handwriting too, partly because of the stresses of time-pressure. Beyond this, undertaking documentation in real-time is highly demanding. Talking to patients, asking them questions, listening carefully, while documenting their responses, is a multitasking nightmare. Also humans don’t tend to multitask - rather, they task toggle - and this invites more opportunities to make mistakes.

AI alleviates some of these problems but of course introduces new sources of error. But it is just plain wrong to assume that sans AI our records are error free. Whether our documentation is now better or worse is an issue which requires more evidence.

It is also worth noting that the rise of ambient AI has coincided with patients gaining much greater access to their medical records, which may make us judge AI-generated errors more harshly. Since October 2023, GP practices in England have been required to offer patients online access to new information added to their records.

As ambient AI has been adopted over the same period, patients may now be spotting mistakes in AI-generated notes that they would previously never have seen in clinician-written records. This creates a risk of giving doctors a pass for errors that were always there but largely hidden from view.

We also tend to hold AI to higher standards than humans. AI is shining a light on the fact errors occur in our medical records, and it is important that patients are fully engaged to spot them. This is also because some errors only patients will be able to identify. There is certainly nobody more invested in the accuracy of the record than the patient.

And this is where Healthwatch is surely right: there needs to be a simple mechanism for patients to flag mistakes and get them corrected. The challenge is doing this in a way that doctors are able to respond, and recognise they need to do so, without simply adding another administrative burden.

AI scribes are spreading quickly. NHS England currently lists 27 suppliers on its ambient voice technology registry. So yes, scrutinise the error but do not assume that the benchmark was, and is, human perfection.

Charlotte Blease wants smarter healthcare for patients.

Author of Dr Bot: Why Doctors Can Fail Us and How AI Could Save Lives (Yale University Press, 2025)

The second oldest anatomical theatre in the world - the Gustavianum in Uppsala University, Sweden (1662–1663) (the oldest is at the University of Padua).

In Europe there is surprisingly scarce research into clinicians’ use of generative AI tools. Beyond, this we are aware of no country that has investigated the adoption practices of both doctors and nurses’. Our latest research bucks the trend. In a survey of almost 5,900 Swedish physicians and nurses conducted in 2025, more than one in five doctors and one in ten nurses said they had purposefully used generative AI to support their clinical work.

Similar to our findings on psychotherapists’ use of AI, we found that ChatGPT overwhelmingly dominated: 84% of Swedish physician users and 81% of nurse users reported using it. Naturally, this matters because ChatGPT is not a specialist medical system, it is a general-purpose commercial chatbot with a buyer beware for healthcare.

Swedish clinicians were not merely using AI for documentation purposes - a leading use in UK surveys. Instead, more than half of the doctors surveyed said they used generative AI to assist with differential diagnosis.

With my PhD student Carolina Garcia who is first author of our Swedish survey. Pictured in Rome during a conference in digital mental health.

Only a small minority - 6.2% of doctors and 3.9% of nurses - reported that their employer provided training in generative AI during the previous year. Moreover, outright prohibition was reportedly rare: just 1.5% of physicians and 0.7% of nurses reported that their employer had prohibited GenAI use.

In my team’s previous UK studies, we found much the same thing among GPs: doctors were experimenting with generative AI with formal training and guidance lagging behind. Healthcare organisations understandably move cautiously. Unfortunately, consumer technology does not wait for committees.

While doctors were more than twice as likely as nurses to report using GenAI, younger clinicians were more likely to use it than older ones, and adoption also varied according to employer type. Clinicians working for private providers were more likely to say they adopted these tools.

Beautiful Uppsala in winter. Uppsala University is where the research team is based.

This was a cross-sectional study, and - as a survey - it was self-reported. The response rate was around one third (which is pretty good for a survey), but our respondents differed somewhat from non-respondents, particularly by age. So, the study should be read with the usual caveats.

Because academic journals take around a year to peer review and publish papers, the paper is currently a preprint - in other words, it is not yet peer reviewed. Better to report it now, than wait for it to be an outdated historical artifact.

Still, the point remains: in Sweden, doctors and nurses are using commercial AI for core clinical activities.

Charlotte Blease wants smarter healthcare for patients.

Author of Dr Bot: Why Doctors Can Fail Us and How AI Could Save Lives (Yale University Press, 2025)

Image from https://www.magnific.com/

This piece draws on my new open-access paper, “AI as a wiser friend? Epistemic risk, companionship and the ethics of generative AI design,” published on 11th August 2026 in the Journal of Medical Ethics: Practical Bioethics.

Generative AI chatbots have gained a reputation for being people-pleasing. It is now commonplac…

Read more

Image by https://www.magnific.com/

If a doctor recommends a drug, we tend to expect some version of the following discussing: what the drug is called, what it is for, how well it works, what are the risks, the alternatives, the costs, and what might happen if we decline to take it. There is a veritable literature on how rushed these conversations can be,…

Read more

Image by Magnific.com

Somewhere today, a therapist will type details of a client’s problems into ChatGPT and ask what to do next.

In our new, global survey of 766 mental health professionals across 30 countries, 55 per cent said they had deliberately used consumer generative AI tools in psychotherapy practice. Among those users, 85 per cent told us they u…

Read more

Image by magnific

The fourth estate voices mounting unease over chatbot-induced harms, including for people with mental health conditions. Despite this growing angst, we still know little about who - or what - on balance, better helps patients. There is still limited research on this issue. In this week’s Substack I examine some of the findings on humans…

Read more

An artist in residence at the Cologne conference sketched these mini vignettes

Conference season continues, and last week I found myself making essentially the same argument twice, to different mental health professionals at two different events. The first was to a room full of psychiatrists at the University of Oxford, and then to around 300 Rogerian ps…

Read more

For five years in Boston, that book existed mostly in my head, or on my laptop. It was written before work, after work, at weekends, and during those suspended months of COVID. It was an itch I had to scratch. I rarely spoke about it o…

Read more

But our survey also asked a different question: what do GPs think AI will mean for the future of their profession? The answers reveal neither complacency nor panic. Curiously, however, they paint a picture of a profession expecting significant …

Read more

Image by vectorjuice on Magnific

Last week, a remarkable new Nature paper introduced MIRA (Medical Intelligence for Reasoning and Action). This AI agent is designed not just to answer medical questions but to carry out many of the tasks doctors perform every day. Developed by Jakob Nikolas Kather, Dyke F…

Read more

Read more

And now to the topic du jour - or rather, of May 26th.

Tony Blair’s comments on artificial intelligence in an essay, published on that date, caught my attention. Why? Because he is on…

Read more

With Cory Doctorow

Let me begin by saying, I am not the festival type. My idea of personal hell is Glastonbury. So, when I spotted the tents, the tie-dyed dungarees, and the strong smell of incense, an inner dread set in. Yet, in spite of myself, or perhaps due to low expectations, I rather enjoyed the book festival called “How The Light Gets In” in Hay-…

Read more

Our first meeting in London.

I knew nothing of Richard Lehman until he revealed himself to me. First he contacted…

Read more

Read more

The debate about generative AI in healthcare has acquired a slightly surreal, Monty Python-esque quality. Policymakers continue discussing these tools as though they are approaching steadily and visibly over the horizon. Meanwhile, doctors are already using ChatGPT to draft letters, summarise notes, and even to assist with core clinical skills like diag…

Read more

I grew up in Belfast, in a place where borders or “peace lines” were not only contours on a map. They were perceived in schools, workplaces, and picked up on in social conversations that implied which “side” you were on. Even now, stereotypes still subtly shapes how people think, what they assume, what they don’t say.

Recently reignited debates about a u…

Read more

Image by Freepik

Digital therapeutics are everywhere now—apps for depression, AI chatbots for anxiety, tools promising to improve sleep, focus, even chronic illness. There are thousands of them available to download. These apps are increasingly marketed not merely as well-being interventions or lifestyle add-ons but as treatments. And yet a basic issue r…

Read more

The conventional wisdom is we should ideally have the same physician from our thirties into our seventies - someone who remembers our medical history, family circumstances, and perhaps even our tendency to catastrophise mild symptoms, or, conversely, to…

Read more

Read more