Two letters published in response to an NHS watchdog's warning about AI medical scribes go after the one claim the technology is sold on: that it saves clinicians time. Mary Gibbs, a GP working out-of-hours shifts in Manchester, argues that AI-generated consultation records are long, repetitive and sometimes self-contradictory, and that the checking they force on the doctor who produces them and the doctor who later reads them adds up to more work, not less. The report she was answering, published on 31 August, said AI secretaries used by doctors were recording medication names and diagnoses wrongly.
Gibbs's complaint is procedural rather than technical, which is what gives it weight. She acknowledges the AI has several problems, above all a tendency to mishear what was said and to contradict itself, and that this can be dangerous for patients. She chose instead to write about the time.
Many of the patients she sees have already been assessed over the phone by another clinician. Before calling a patient in, she reads the record of that first conversation. When AI was used to produce it, she wrote, her "heart sinks": the consultation is almost always long, repetitive, and sometimes contradictory.
Worse, the history she takes herself often differs noticeably from the one in the record, which leaves her unable to treat the text as reliable. Substantial errors, she wrote, turn up far more often in AI-assisted consultations than in notes colleagues typed by hand.
The obvious rejoinder is to blame the first clinician for not proofreading. Gibbs gets there first: anyone who has tried to check their own writing, particularly with no time to spare, knows how hard it is. Her question is where the saving is supposed to come from, if the first doctor has to read back a long, repetitive description and the second cannot trust it and has to verify it again.
She adds a second loss, slower and harder to see. Turning a patient's often complicated story into a clear description another clinician can grasp quickly is a skill, and one that takes time to acquire — any medical student would confirm it. If it stops being learned and maintained, she argues, both doctors and patients lose.
The second letter is a smaller story with a sharper edge. Debbie Cameron, writing from Formby, Merseyside, worked as a secretary for a consultant who trialled a speech recognition system for a week. It was dropped after "I recommend taking lansoprazole for two weeks" came out as a proposal for a two-week holiday in Lanzarote. The patient, Cameron notes, would have been delighted.
Read together, the letters describe a measurement problem more than a technology problem. The time saving is claimed at the point where it is cheapest to observe: the clinician who was talking anyway no longer types. The cost turns up one desk over, in a second clinician's re-check, and nobody is counting that side of the ledger. This reads like a classic productivity illusion — the work has not been removed, it has been relocated to a place where it does not register as work.
Cameron's anecdote points at something Gibbs implies but does not spell out: the error class has changed. A typo looks like a typo. Lanzarote looks like a word somebody meant. Hand-typed notes fail in ways a reader's eye catches; transcription failures produce fluent, plausible text that can only be caught by checking it against the patient. That is a different shape of risk from the one clinical records were built to absorb.
The more interesting question is the one neither letter can answer and nobody appears to have asked: what is the net time cost across the whole chain? Gibbs is describing a two-clinician workflow and a document read at least twice. A saving measured on the first pass and a burden imposed on the second is not a saving, and until someone times the second reader, the claim is an assertion.
Notably absent from any of this is a suggestion that the scribes get better at the part that matters. Accuracy on drug names is a solvable engineering problem, and it will probably be solved. Writing a handover note another doctor can absorb in thirty seconds is a different problem, and it is the one Gibbs says trainees will stop practising. The doctors who follow her stand to inherit both at once: records that read well and cannot be trusted, and no habit of writing the kind that can.