TL;DR

Roughly two in five UK GPs now let an AI tool listen to consultations and write the notes. A peer-reviewed review from the University of Edinburgh, published in the BMJ’s digital health and AI title, finds those tools reliably capture what was said and routinely lose how it was said. The researchers also warn of “cognitive offloading” — clinicians who no longer recognise their own notes, or recall the patient at the next appointment.

What the review found

Ambient scribes transcribe the doctor-patient conversation and generate notes and letters from it. The Edinburgh team credits them with real gains: less paperwork, more attention available for complex work. Documented benefits elsewhere are substantial — a Dudley clinic using the Heidi tool cut a patient-letter backlog from six months to a fortnight.

The losses are subtler. Facial expression, gesture and emotional state do not survive transcription. Summaries lean towards clinical fact at the cost of the patient’s own account of their illness. And patients who know a machine is recording become more guarded about substance misuse, domestic abuse and mental health — precisely the disclosures that are hardest to elicit and most consequential to miss.

The deskilling problem

The finding professionals should sit with is about the clinician, not the software. Writing notes by hand is not merely administrative; it is where reasoning and reflection happen. Hand that step to a machine and, the review found, memory recall and skill development weaken with it.

“Many clinicians are excited about ambient AI scribes, because they promise to cut down on paperwork,” said Dr Lucas Seuren, one of the Edinburgh researchers. “But the experiences of patients are poorly considered, and there are real risks that the patients’ stories are lost. This can further disadvantage people who already face marginalisation in health and social care services.”

Looking forward

This lands in a week when Resultsense has already covered GPs carrying unlimited personal liability as AI moves NHS data, and only 14% of England’s teachers feeling confident about AI. The pattern repeats: adoption first, evidence second. The researchers want more work on medium and long-term use, and flag a specific risk for tools built for one health system and deployed in another — a direct caution for NHS buyers procuring products designed around American clinical workflows. Their design principle is worth carrying into any procurement conversation: whatever the tool records, the person’s own account of their illness has to survive it intact.