The failure lands in the two seconds a clinician spends scanning a summary before signing it. A fluent, plausible note is exactly what these systems produce, which is why a wrong one survives that glance [12]. The cases the watchdog collected involved a real sentence coming out slightly wrong, not an invented fact [13], and slightly wrong is enough when the sentence names a drug.
The other examples make the pattern legible. One summary left out a consultant's instruction that the patient should seek a repeat prescription for migraine medication [5]. Another recorded a doctor telling a patient to continue their Prozac when that doctor had neither prescribed it nor discussed it [6]. Counting the drug swap, three of the four documented errors turn on medication [1]. Each one is the kind of error a demo would miss, because on the page it doesn't look like a defect.
The scribe drafts, the clinician reviews, and the record is assumed safe, but the watchdog found otherwise: "These inaccuracies may persist in their records if the patient doesn't catch them" [7]. The working verification loop is a person reading their own letter at home, which is the one reader with no way of knowing what the consultation established. Rachel Power, chief executive of the Patients Association, is among those raising it, alongside clinicians including the London GP Shier Ziser Dawood and Charlotte Blease of Uppsala University in Sweden [14]. The objection is less to the technology than to its arrival without a safety net [15].
The regulatory mechanism explains how that happened. MHRA guidance published in August sets the line: a system that only transcribes what was said is not a medical device, while one that suggests a diagnosis or a treatment may be, which loads enormous weight onto how a vendor describes its own product [9]. A scribe sold as passive transcription skips a process a clinical assistant would have to complete [10]. Meanwhile documentation is the administrative burden doctors complain about most, and anything that reliably removes an hour of typing a day gets adopted whether or not it has been assessed [11].
For anyone rolling out a summarizer into a system of record, the useful grid has two axes. Does the output become durable and authoritative once written, and does a human who knows the ground truth read it before it commits? Notes that go nowhere and get checked are cheap either way. The dangerous quadrant is durable and unchecked, and NHS scribes sit in it: the artefact is the clinical record and the letter to the patient [3], and the only reliable reader is downstream of the truth.
The fix comes with a cost, and it's worth being direct about it. Assign the verification read to someone who has the source material in front of them, and accept that this gives back part of the hour you bought. The watchdog's own remedy, telling patients a scribe was used and handing them the notes to check [16], turns an accidental safeguard into a designed one, but it does not move the work to someone better placed to do it.
A buyer can settle this before signing. Ask what share of notes are corrected before sign-off and who did the correcting. Time saved per clinic is the number the vendor will bring to the meeting.
