Ontario auditors find AI medical scribes routinely fabricate clinical findings

Ontario’s healthcare regulator has uncovered significant accuracy failures in AI medical-scribing tools approved for use by physicians and nurse practitioners. According to The Register AI, the Office of the Auditor General of Ontario evaluated 20 vendor systems and found that 9 reportedly fabricated clinical findings—such as ruling out masses or noting patient anxiety—that were never discussed in the original patient-clinician recordings. The audit raises urgent questions about how AI tools are vetted before deployment in clinical settings where inaccurate documentation can directly affect patient care.

AI Systems Inserting Wrong Drug Information and Missing Mental Health Details

The audit’s scope extended beyond fabrication. According to The Register AI’s report on the findings, 12 of the 20 evaluated systems inserted incorrect pharmaceutical information into patient notes, while 17 systems “missed key details about the patients’ mental health issues” that were discussed in the original recordings. The source notes that six systems either omitted mental health issues entirely or captured only partial information. These gaps are particularly concerning because mental health history is central to holistic care planning and medication interactions.

Flawed Procurement Weighting Prioritized Business Presence Over Accuracy

The audit identifies a structural problem in how Ontario evaluated vendor proposals. According to The Register AI, the evaluation framework assigned only 4 percent weight to the accuracy of medical notes, while 30 percent of the total score hinged on whether a vendor had domestic operations in Ontario. This weighting structure—which prioritizes business and regulatory factors over clinical performance—suggests that procurement officials may not have fully aligned evaluation criteria with patient-safety outcomes.

OntarioMD, the physician-support organization involved in the procurement, recommended that doctors manually review AI-generated notes for accuracy, yet The Register AI reports that none of the approved AI Scribe systems include a mandatory attestation feature that would enforce such review workflows.

Why This Matters

Healthcare organizations across Canada and beyond that are considering AI-assisted documentation tools should treat this audit as a red flag for procurement design. The Ontario audit demonstrates that vendor evaluation frameworks weighted toward business criteria (domestic presence, implementation cost, feature breadth) can produce approvals that fail on foundational clinical safety metrics. Regulators and health authorities renewing or expanding AI Scribe contracts should establish minimum accuracy thresholds—such as ≥90% correctness on drug names, patient-reported symptoms, and clinical findings—before rollout. The absence of mandatory attestation in approved systems is itself a gap: any AI Scribe deployment should require clinician sign-off workflows that are baked into the system interface, not left to voluntary compliance.