Doctors in North Savo warn of patient safety risks from AI speech recognition errors

Doctors in Finland’s North Savo wellbeing services region report that AI-powered speech recognition software is generating dangerous errors in patient records, including incorrect medication dosages and misinterpreted diagnoses, according to a Yle investigation.

General practitioner Keijo Eskola described a near-miss where the system transcribed a prescribed 25-milligram dose as “2, 0, 5 milligrams”—a misreading that could have been fatal, depending on the drug. In another case, the AI omitted the word “no” from a dictation, transforming “the patient does not have a brain tumour” into “the patient has a brain tumour.” The software also struggles with specialised medical terms, names, and addresses.

The region’s wellbeing services authority confirmed five error reports linked to the system in the past six months, though none resulted in serious harm. “No severe patient safety incidents have occurred,” said medical director Antti Hedman, acknowledging the flaws. The authority laid off dozens of transcription staff earlier this year, shifting responsibility for accuracy checks to doctors—who remain legally liable for the text.

Eskola warned that the added workload and stress increase the risk of further mistakes. Anu Muraja, a specialist physician and board member of the Finnish Medical Association, called for urgent fixes, noting that while minor dictation errors have always existed, “when errors start threatening patient safety, the system must be corrected.”

The region uses Inscripta Direct, an AI tool designed for real-time healthcare documentation, which the manufacturer claims achieves 96–98% accuracy. Muraja also criticised government proposals to allow AI to assess patient treatment needs, arguing that human clinicians’ ability to weigh options and ask follow-up questions cannot yet be replicated by machines.