AI scribes confuse medicines and diagnoses in the NHS

What this is about

AI scribes listen during medical consultations and automatically create transcripts, summaries or entries for patient records. Healthwatch England, the statutory patient champion for England's health service, is now warning about specific errors: one system confused diagnoses, another named the wrong medicine, and a third omitted an important instruction about obtaining a repeat prescription.

The cases were publicly reported on August 31, 2026. They matter because clinicians in England are already using 27 different AI scribes and the UK government wants broader adoption. In several cases, patients noticed errors that medical staff had not caught.

What AI scribes actually do

An ambient scribe records the consultation or processes its audio. Speech recognition turns statements into text. A language model then organises symptoms, diagnoses, medicines and next steps into a structured note.

This can reduce writing and help clinicians pay more attention to the person in front of them. But the technology cannot reliably determine which of two similar-sounding medical phrases was intended. In one documented case, “null demyelination” became “demyelination.” A statement excluding serious nerve damage therefore appeared to become a diagnosis.

In another case, the system replaced a prescribed drug with a medicine whose name sounded similar. A separate summary letter omitted an instruction telling a patient to ask her GP for a repeat migraine prescription.

Why it matters

An error in a patient record can outlive a mistake in an ordinary conversation. Later doctors, pharmacists or hospitals may copy the incorrect information. A false diagnosis can affect tests, insurance questions and the patient's psychological wellbeing. A wrong drug name can create immediate danger.

NHS England's guidance therefore calls for human review, clear accountability, consent and secure data handling. The cases reveal an attention problem in practice: the more fluent and professional an automated note sounds, the easier it may be to assume that it is correct.

There is also a regulatory dispute. Healthwatch has criticised the decision by the UK medicines and medical devices regulator not to classify these scribes as medical devices as a group. That leaves no uniform nationwide safety review covering the whole product category.

In plain language

An AI scribe is like someone taking down a recipe during a fast cooking lesson. Most ingredients and steps may be recorded correctly. But if “no salt” becomes “salt,” the note can look tidy while producing a completely different result. The cook must therefore check the recipe before sharing it.

A practical example

A GP practice handles 60 consultations per day and uses automated summaries for all of them. If just one percent of notes contained a relevant error, that would amount to three flawed records per five-day week.

For one patient, the system mishears the name of a blood-pressure medicine and inserts a similar-sounding drug. The doctor skims the note and approves it. Two weeks later, a covering doctor reads the entry and assumes the incorrect medicine was genuinely prescribed. A robust review routine would therefore compare medicines, doses, allergies, diagnoses and next steps individually against the conversation or existing record.

Scope and limits

• The reported cases demonstrate real risks but do not provide a reliable error rate across all 27 systems in use.
• Humans also make documentation mistakes. Without comparative studies, it is not possible to say whether AI scribes are more or less error-prone overall.
• Performance may change substantially with accents, background noise, multiple speakers and complex medical histories.
• Automated notes must not replace supervised medical decisions. Providers need clear correction routes for patients and traceable accountability.

SEO & GEO keywords

AI scribes, NHS, Healthwatch England, patient record, medication error, diagnosis error, ambient scribe, medical documentation, patient safety, speech recognition, clinical AI