Ontario audit: AI medical scribes routinely garble prescriptions and patient facts

Ontario auditors evaluating AI Scribe tools used by physicians found pervasive accuracy failures, with 60% of the systems reviewed mixing up prescribed medications in generated patient notes. The errors weren’t subtle edge cases — they involved basic clinical facts that downstream care depends on, raising direct patient-safety concerns about tools already being deployed in exam rooms.

The findings undercut vendor claims that ambient-listening LLM transcription is ready to offload clinical documentation from overworked doctors. Drug-name confusion in particular is the kind of hallucination that can flow straight into prescriptions, pharmacy systems, and follow-up visits before anyone catches it, and it points to the limits of speech-to-structured-note pipelines that lack robust grounding against a patient’s actual chart and formulary.

For health systems, the audit is a prompt to treat AI scribes as draft generators requiring physician verification rather than autonomous documentation, and to demand measurable accuracy benchmarks — especially around medications, dosages, and allergies — before procurement. It also strengthens the case for regulators to set minimum evaluation standards for clinical AI tools that are currently shipping with little independent oversight.