AI Systems for Ontario Doctors Hallucinate: Auditor General

Why It Matters

Inaccurate AI‑generated notes jeopardize patient safety and expose Ontario to legal and privacy risks, while unchecked AI use threatens data security across the public sector.

Key Takeaways

• Nine of twenty AI scribes produced fabricated clinical information.
• Twelve systems recorded incorrect prescriptions; seventeen missed mental‑health details.
• Unapproved AI sites made up 94% of staff usage, unsafe.
• Only 3% of 55,000 employees completed mandatory AI‑risk training.
• Auditor recommends tighter security, bias testing, and mandatory note sign‑off.

Pulse Analysis

AI‑driven transcription tools promise to slash clinicians’ paperwork time, with OntarioMD reporting up to 90% reduction. Yet the auditor general’s May 12 audit shows the technology is still error‑prone. Nine of the twenty evaluated scribe systems fabricated patient information that never occurred, and twelve mis‑recorded prescribed drugs. Missed mental‑health cues appeared in seventeen systems, raising the specter of incomplete treatment plans. Such hallucinations, even if confined to testing phases, erode clinician trust and could translate into harmful clinical decisions if deployed at scale.

Beyond health care, the Ontario Public Service’s AI governance appears fragmented. The audit found that 94% of AI interactions by staff occurred on unapproved generative‑AI websites, with 60% flagged as unsafe and 15% hosting inappropriate content. Only six percent of usage involved the sanctioned Microsoft Copilot Chat, which operates within a Canadian‑data enclave. Compounding the risk, a mandatory Responsible Use of AI course reached merely 3% of the 55,000 workforce, leaving most employees unaware of bias, misinformation, and data‑leakage threats.

The auditor’s nine recommendations aim to tighten security, enforce bias testing, and embed mandatory sign‑off controls on AI‑generated notes. If implemented, these measures could restore confidence in digital health tools and reduce the province’s exposure to privacy breaches and liability claims. Other jurisdictions are watching Ontario’s experience as a cautionary tale, prompting a broader conversation about AI policy, procurement standards, and environmental impact assessments across Canadian governments. Robust oversight will be essential for leveraging AI’s efficiency gains without compromising patient safety or public‑sector data integrity. The province’s next steps will likely set a benchmark for AI accountability nationwide.