Back to Articles
AI Transcriptions Are No Time-Saving Measure for Doctors

The Guardian

ENRICHED

Details

Date Published
7 Sept 2026
Priority Score
2
Australian
No
Created
7 Sept 2026, 08:00 pm

Authors (2)

Description

Letters: Dr Mary Gibbs says there are a number of problems with AI, particularly its tendency to misunderstand, as Debbie Cameron can attest

Summary

This correspondence highlights critical reliability failures in AI-driven medical transcription, specifically noting how 'hallucinations' and errors in drug names or diagnoses create significant patient safety risks. The authors argue that AI tools currently increase the cognitive load on clinicians due to the need for rigorous proofreading and the erosion of essential medical synthesising skills. While the content focuses on immediate clinical hazards rather than existential AI risks, it underscores the systemic dangers of deploying frontier AI systems in high-stakes environments without robust accuracy and verification frameworks.

Body

‘Turning an often complex history by the patient into a clear account that a colleague can assimilate quickly is a difficult skill to learn.’ Photograph: Getty/iStockphotoView image in fullscreen‘Turning an often complex history by the patient into a clear account that a colleague can assimilate quickly is a difficult skill to learn.’ Photograph: Getty/iStockphotoLettersAI transcriptions are no time-saving measure for doctorsDr Mary Gibbs says there are a number of problems with AI, particularly its tendency to misunderstand, as Debbie Cameron can attestI am an out-of-hours GP, and I am glad to see a news report on this issue, which has been causing me increasing concern at work (Doctors’ AI scribes get names of drugs and diagnoses wrong, NHS watchdog warns, 31 August). There are a number of problems with AI, particularly its tendency to misunderstand and contradict itself, which can lead to dangers for the patient, but I will concentrate on the supposed “time-saving” aspect.Many of the patients I see will have been assessed by someone else first on the phone, and I am used to reading through a triage assessment before I call in the patient. My heart sinks when I see that AI has been used – the consultation is invariably long, with duplications, and sometimes contradictions. Often, any history I take turns out to be significantly different from what is listed. So I can never trust that it is correct – I can state with certainty that I find meaningful errors far more often in AI consultations than in those typed out by colleagues.You may blame the original clinician, but anyone who has tried to proofread something they have created, particularly under time pressure, will know how difficult this is. So how can this save time, when the first clinician has to check through a long, repetitive account, and the second clinician can’t trust what is written and has to check it again?In addition, turning an often complex history by the patient into a clear account that a colleague can assimilate quickly is a difficult skill to learn, as any medical student can tell you. If we do not learn and maintain this skill, we will lose a lot, and so will our patients.Dr Mary GibbsManchester Re the risks of AI diagnosis and medication being mistranscribed, I have first-hand experience of the problems. A consultant I worked for as a secretary had a week’s trial of voice recognition software. He gave up the idea when “I recommend two weeks on lansoprazole” turned into suggesting a two-week holiday on Lanzarote. A prescription I’m sure the patient would have been more than happy to try.Debbie CameronFormby, MerseysideExplore more on these topicsDoctorsNHSAI (artificial intelligence)HealthComputinglettersShareReuse this content