Loading today’s stories
Leadership1 distinct publisher3 min readPublished
The scribe errors Healthwatch documented reached patients' records before anyone queried them. In two of the cases, it was the patient who did the querying. England has 27 products in use and no device classification.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
The demyelination case turns on one word, and that word makes this a governance problem, not a transcription one. The hospital record should have read "null demyelination", and the negation carries the entire clinical meaning [2]. A dropped negative does not read as a malfunction. The summary stays fluent clinical prose, which is the form a reviewer working at speed is least likely to stop on. The patient, herself an NHS health professional, queried the tool's version of her MRI result and the hospital corrected it [2]. She described being given an incorrect diagnosis because of AI and then being told it was a typo [3].
Detection here came down to incentive, not expertise. Healthwatch England, the statutory NHS patient champion, says it has heard multiple stories from patients who noticed errors when a health professional had not, and that inaccuracies may persist in records if the patient does not catch them [5]. In the second documented case an AI scribe confused the prescribed drug with a different one of similar name, and the patient rather than the doctor identified it [4]. Across the three cases described, the patient is the named detector in two [1], and the failure modes differ: two inserted wrong content, one omitted an instruction the consultant had given [2]. Reading a summary for plausibility catches neither reliably.
The fair objection comes from inside the evidence. Dr Charlotte Blease of Uppsala university, who studies AI use in healthcare, says AI can and does make mistakes, and adds that doctors can and do make mistakes too [12][14]. Both halves are true. What differs is the handling, not the error rate. A clinician's mistake has an author and a complaints route. A scribe's mistake sits across the 27 different products already in use by GPs and hospital doctors in England [9] with no England-wide standard for what a review consists of, because the MHRA has decided not to classify these tools as medical devices [8]. Healthwatch's own request is procedural, asking for clarity over how patients can report and get corrected any mistakes made by the tools or by the professionals using them [6].
The productivity case is where the missing step gets expensive. The government's 10-year health plan expects scribes to liberate staff from their burden of bureaucracy and administration [10], while Dr Shier Ziser Dawood, a London GP writing in the British Journal of General Practice, reported that the need to review every transcript means the tools are not yet saving time [17], and that GPs may be expected to see two more patients a day in consultations already among the shortest in the world [18]. Her own scribe recorded her telling a patient to continue their Prozac, a drug she had neither prescribed nor discussed [16]. A rollout that books the saving before it funds the check has decided, in practice, that patients will do the checking.
Rotherham shows where that lands. Patients there complained to their local Healthwatch that an AI receptionist used by some GP practices could not understand their strong Yorkshire accents [15], which made the patient-facing watchdog the defect tracker by default. The cheapest correction available this quarter is a governance line, not a procurement one: a named reviewer for each summary before it enters the record, and a logged route back out when a patient reports an error. Deferring it converts an error rate into a records problem later, at the point where ministers have already been warned that the NHS and its medics could be sued over mistakes made by AI [11].
Ranked by verification strength, evidence, and original report placement.
Dr Charlotte Blease, an expert in AI use in healthcare at Uppsala university in Sweden, said: "AI can and does make mistakes."
Blease's research found that GPs who use ambient voice technology believe errors are more likely to creep in when the consultation is with more than one person, with patients with a complex medical history, and with those whose first language is not English.
Blease added: "The fact is, doctors can and do make mistakes".
Healthwatch England, an NHS watchdog, warned that AI technology transcribing patients' consultations with doctors can put them at risk by getting the names of drugs and illnesses wrong, and said that unless detected such errors could end up in patients' medical records and affect their care.
In one case an AI scribe's summary wrongly said a woman had demyelination, serious nerve damage that can lead to multiple sclerosis; only when the patient, an NHS health professional, queried the tool's record of her MRI scan result did the hospital correct it to what it should have been, "null demyelination".
The woman, who asked not to be named, said: "This was eventually corrected but was a very traumatising experience to be given an incorrect diagnosis because of AI and then be told it's a typo."
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 30, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
A blood-draw robot clears the FDA at 1m49s a stick. The missing number is per supervisor1 distinct publisher
invest
Wispr's $280M bet: voice input's problem is accuracy, not appetite1 distinct publisher
science
Webb splits Hephaistos's two newest Dyson sphere candidates into M dwarfs and background galaxies1 distinct publisher
invest
A $2B bet that voice replaces the text box, not just the keyboard3 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Three vivid cases, no denominator
The cases are specific enough to be checkable in kind — 'null demyelination' becoming a diagnosis is not the sort of detail a reporter invents — and they come with a statutory watchdog, a named London GP writing in a medical journal, and an Uppsala researcher. What is absent is arithmetic. Healthwatch's 'multiple stories' is never counted, its findings are quoted rather than published, the 27-product figure arrives without a register behind it, and the litigation warning to ministers has no author. One newsroom is doing all of the reporting.
27 products in, nothing counted
Deployment is real and broad rather than piloted: 27 distinct scribes across GP practices and hospitals, a national plan that makes them central to the analogue-to-digital shift, and a survey that found 1,003 GPs with enough experience of ambient notes to compare them with their own. The ceiling on this score is that no one here counts practices, consultations or vendors, so the depth of use inside those 27 products is unknown.
The promised hours are the overclaim
The overstatement here belongs to the benefit, not the risk. A 10-year plan promises staff liberated from administration; a GP using the tools says she must read every transcript and has therefore saved no time, and the freed minutes are already being converted into an expectation of two more patients a day that no NHS document in this reporting sets. Pushing the other way, the harm side is anecdotal too — three cases and an uncounted 'multiple' — and the paper's own survey line, where most GPs rate AI notes above their own, keeps this from being a straightforward case of a technology sold beyond its evidence.
Four interests speaking, the sellers silent
Every voice quoted has a stake worth naming. Healthwatch and the Patients Association exist to escalate patient grievance, and this is their material. The government has bet a ten-year plan on the digital shift and needs scribes to work. The MHRA's decision not to classify these tools as devices spares it an oversight burden and is reported only through the party unhappy about it. The clinicians' interest is the most concrete of all: whoever signs the record carries the liability for what the machine wrote. Missing entirely are the 27 companies selling the products.
Credible account, single voice
We would stand behind the failure modes and the regulatory gap; we would not yet stand behind any statement about how often this happens. One publisher, no primary Healthwatch document, and no reply from the MHRA, the health department or a single vendor. What lifts our confidence above the floor is that The Guardian prints the evidence against its own framing — Blease on human error rates and the survey in which GPs prefer the machine's notes — which is not the behaviour of a piece straining for a verdict.