May 19, 2026 AI systems approved for use by healthcare providers in Ontario are producing unreliable medical notes, including hallucinated content and missing key clinical details, according to a provincial audit. The review found that nine out of 20 approved systems fabricated information or suggested treatments that were never discussed between patients and clinicians.
The findings come from the Office of the Auditor General of Ontario as part of a broader report examining how artificial intelligence is being used across public services in the province. At the center of the issue is the AI Scribe program, an initiative launched by the Ontario Ministry of Health to help physicians, nurse practitioners and other healthcare professionals automatically generate clinical notes from patient interactions.
To evaluate the systems, officials used simulated doctor-patient recordings. Medical professionals then compared the original conversations with the AI-generated notes to assess accuracy. The results raised serious concerns about the reliability of these tools in clinical settings.
In multiple cases, AI systems inserted information that was never mentioned during consultations. Evaluators reported examples where systems noted the absence of medical issues, such as masses, or described patients as anxious, despite no such details appearing in the recordings. These kinds of inaccuracies, especially in healthcare, could have significant consequences if not caught.
The problems extended beyond hallucinations. Twelve of the 20 systems added incorrect drug-related information into patient records, while 17 failed to capture important details about patients’ mental health. In six cases, systems either partially or completely missed mental health issues that had been clearly discussed during the consultation.
Despite these shortcomings, the tools have already seen adoption across the province. More than 5,000 physicians are currently participating in the AI Scribe program. While there have been no reported cases of patient harm so far, the audit suggests that the margin for error is concerningly high.
Part of the issue, the report argues, lies in how these systems were evaluated and approved in the first place. The scoring system used during procurement placed relatively little emphasis on clinical accuracy. Only 4% of a system’s total score was tied to the accuracy of medical notes, while factors like having a domestic presence in Ontario accounted for as much as 30%.
Other critical areas, such as bias controls, risk assessment and privacy protections, were also given minimal weight, each contributing just a small percentage to the overall evaluation score. According to the audit, this imbalance could lead to the selection of vendors whose systems are not sufficiently reliable or secure for handling sensitive health data.
The report warns that such evaluation practices increase the risk of deploying tools that produce inaccurate or biased medical records. In a healthcare environment, where decisions rely heavily on precise and complete information, even small errors can cascade into larger clinical risks.
Organizations involved in the rollout have acknowledged the need for oversight. OntarioMD, which supports physicians in adopting new technologies, has advised doctors to manually review AI-generated notes before using them. However, the audit highlights a critical gap: none of the approved systems include a mandatory attestation feature requiring clinicians to confirm the accuracy of the notes.
This means the responsibility for catching errors rests entirely with already busy healthcare professionals, without system-level safeguards to enforce verification.
