An Ontario audit has uncovered serious accuracy problems with AI-powered clinical documentation systems used by physicians. The investigation found that artificial intelligence notetakers regularly generate fabricated medical information, including invented therapy referrals and incorrect prescription details that never occurred during patient visits.
The audit examined how these systems transcribe and summarize doctor-patient interactions. Instead of accurately recording what transpired, the AI frequently hallucinates clinical facts. Therapy referrals appear in patient records despite never being discussed. Medication recommendations include drugs the physician never mentioned or doses that differ from what was actually prescribed.
This matters because medical records form the foundation of patient care across multiple providers. When emergency rooms, specialists, and pharmacies access these notes, they rely on their accuracy. Fabricated referrals can lead patients to pursue unnecessary treatments. Incorrect prescriptions create dangerous drug interactions or dosing errors. Wrong information in records also complicates future diagnosis and treatment planning.
The problem stems from how these systems work. Large language models underlying AI notetakers excel at pattern matching and generating plausible-sounding text, but they don't truly understand clinical context or medical knowledge. They occasionally confabulate details to fill gaps in what they heard, especially in noisy clinical environments or when terminology is ambiguous.
Ontario's healthcare regulators did not name specific vendors in the audit findings, but similar systems from companies like Ambient and Microsoft are widely deployed in North American clinics. These tools promise to reduce physician administrative burden and improve note quality. Instead, they're introducing a new failure mode into healthcare systems.
The findings raise urgent questions about AI deployment in medicine. Vendors and healthcare facilities must establish verification workflows where physicians validate every AI-generated note before it enters patient records. Without human review, these systems pose genuine risks to patient safety. The Ontario audit demonstrates that convenience cannot override accuracy in clinical documentation.