How One Missing Word turned an AI Scribe into a Patient Safety Liability
AI scribes are being adopted rapidly throughout the NHS, with the promise to reduce one of clinicians’ biggest administrative burdens: documentation. A new warning from Healthwatch England, however, has exposed how even tiny transcription errors can become serious patient safety problems.
In one case, a patient was told they had demyelination, the nerve damage associated with conditions like multiple sclerosis. They didn’t. The original test result actually read “null demyelination”, but the AI scribe dropped one word, null, that completely reversed the meaning.
Other reported errors included replacing a prescribed medicine with another drug of a similar name, omitting instructions to request a repeat migraine prescription and recording that a patient should continue taking Prozac despite the fact the doctor had neither prescribed nor discussed it.
The problem is not necessarily that these systems produce obviously broken notes. The danger is that AI-generated summaries are often so fluent, structured and plausible, that you might not even notice subtle inaccuracies during a quick clinician review.
Healthwatch warned: “These inaccuracies may persist in their records if the patient doesn’t catch them”. This puts us in the very uncomfortable position of having the patients themselves as the final safeguard against inaccurate record-keeping, despite the fact they’re unlikely to be qualified to know exactly what the clinician intended to record.
The regulatory position adds another layer of complexity. AI scribes that transcribe and summarise are not necessarily classified as medical devices by the MHRA. This allows them to avoid the scrutiny of the regulatory process that a marketed clinical assistant scribe would need to complete. So two extremely similar products could be held to different standards of safety and effectiveness simply based on how their purpose is marketed.
“The remedy the watchdogs point to is unglamorous and probably right. Patients should be told when a scribe is being used and given their notes to check, which turns an accidental safety mechanism into a deliberate one”
Ana-Maria Stanciuc, Editor-in-Chief of TNW
This is yet another symptom of the wider trends of AI in healthcare: adoption moves faster than governance. Clinicians want to save time on documentation, but once an AI-generated note becomes part of the permanent medical record, even a small error can follow a patient through future care.
The answer is unlikely to be abandoning AI scribes entirely. Instead, we must emphasise clearer oversight, reliable verification processes and encourage the adoption of specific AI scribe tools that we know have been properly validated. Because AI does not need to hallucinate an entire diagnosis to cause harm, sometimes all it takes is one missing word.
