What the research found
King's College London researchers recorded a single speaker reading 100 surgical case reports and ran the audio through eight systems: three commercial clinical transcription products, four speech-recognition services and an experimental pipeline that used a large language model to correct errors. Between 30% and 68% of the 800 transcripts contained at least one clinically significant error, and 53 errors carried potential for severe harm or death, concentrated in medication type and dose (39.6%) and investigations and lab results (30.2%). The language-model correction step cut affected transcripts from 53% to 30% but introduced some new errors, and a quiet single-speaker setup is likely to understate real-world error rates.
Ambient scribes are being rolled out for the time they save, but the errors that matter cluster in doses and results, exactly where a rushed sign-off is least likely to catch them.