An AI transcription tool deployed for Ontario physicians hallucinated content and introduced errors into clinical records, according to findings from an official auditor cited by CBC News. The system was operating in live medical documentation workflows at the time the errors were identified.
The auditor's confirmation moves this beyond anecdotal reporting — clinical hallucinations in this case are documented through a formal review process, not self-reported by the vendor or surface-level user complaints. Specific details on error type, volume, or patient impact were not confirmed in the available signal.
The finding adds to a growing body of audited evidence that LLM-based transcription systems can silently corrupt high-stakes records without triggering obvious failure signals. Unlike general productivity contexts, clinical documentation errors can propagate into diagnosis, prescribing, and billing workflows before detection.
For builders and operators deploying AI transcription in regulated environments: the Ontario case reinforces that human review checkpoints are not an optional layer — audited failure evidence now exists to support that requirement in procurement, compliance, and liability conversations.