Artificial intelligence tools now deployed across the NHS to transcribe doctor-patient consultations are generating clinically significant errors in medical records, according to Healthwatch England. The watchdog has documented cases in which the software has altered diagnoses and misrecorded prescribed medications, producing notes that read fluently but contain factual reversals a clinician may not notice on a quick review. Healthwatch has not disclosed how many patients have been affected nationally, and the full scale of the problem across NHS trusts remains unverified.

The examples collected so far are specific. In one case, a negative test result for nerve damage was transcribed as a positive finding of demyelination, inverting the clinical meaning of the consultation. In others, similarly named drugs have been swapped in the record. One patient's notes documented instructions to continue taking Prozac, a medication that had never been discussed or prescribed during the appointment. Follow-up instructions, including reminders to collect repeat prescriptions, have also gone missing from transcribed summaries.

These are not typographical slips. They are errors that change what a patient believes about their own body.

The reason doctors are missing them lies in how the technology writes. AI scribes do not produce garbled or obviously broken text. They produce prose that sounds like a competent clinician wrote it, complete with the expected structure and terminology of a clinical note. A doctor scanning a summary before signing off is looking for something that seems wrong. When the sentence reads naturally, there is nothing to catch the eye. The fluency of the output is precisely what makes the errors invisible.

There is precedent for this pattern of failure. Speech recognition software introduced into radiology departments during the 2000s produced similar substitution errors, misheard drug names and reversed negatives, which studies later found persisted in final reports because radiologists trusted the transcription more than they checked it. The lesson from that period was that confidence in a tool's fluency is not the same as confidence in its accuracy. That lesson has not been applied here.

At present, 27 different AI scribe tools are in use across the health service, with no centralised body overseeing their accuracy or auditing their output. Each has been procured separately, often by individual trusts or practices, with no shared standard for what counts as an acceptable error rate.

The tools have avoided formal regulation through a definitional gap. The Medicines and Healthcare products Regulatory Agency classifies software as a medical device when it contributes to clinical decision-making. Vendors market their scribes as passive transcription tools, simply converting speech to text, rather than systems that shape diagnosis or treatment. That framing keeps the software outside the MHRA's definition of a medical device, and exempts it from the safety and efficacy testing that classification would require before deployment.

Adoption has continued regardless. Clinicians face a documentation burden that regularly consumes hours of each working day, time taken directly from patient care. Software that promises to eliminate that burden is an easy sell to trusts and GP practices under pressure, and the incentive to adopt quickly has outpaced any incentive to test rigorously first.

The result is that patients have become the last line of defence against a system with no formal oversight.

Healthwatch England and other patient groups are now calling for two specific changes. The first is mandatory disclosure: patients should be told explicitly when an AI scribe is recording their consultation, rather than discovering it by inference. The second is a formal review process, in which patients are given a structured opportunity to check their own clinic letters against what was actually said, rather than relying on chance discovery of an error.

Without that check, the record stands as written. A misheard word becomes a diagnosis. A diagnosis becomes a treatment plan.

‍