Ontario's AI auditor finds accuracy gaps, bias risks in health care AI tools

Ontario's AI rollout in healthcare has accelerated faster than oversight, with a provincial audit revealing significant accuracy and governance gaps that raise questions about patient safety as artificial intelligence spreads through the health system.

The C.D. Howe Institute, a think tank that advocates free-market economic policy, held its fourth working group meeting on June 23 to examine the adoption, oversight, and commercialization of artificial intelligence in Canada's healthcare system. The meeting featured an audit by Shelley Spence, Auditor General of Ontario, examining how Ontario procured and evaluated AI tools — particularly AI scribes, which automatically generate clinical documentation from physician-patient interactions.

Across the Ontario Public Service, the audit found that staff could access unsanctioned and unsecured AI tools from government-issued devices. Approved AI tools were used at a rate of roughly 6 per cent, against a 94 per cent usage rate for non-approved generative AI websites. Only about 3 per cent of staff had completed AI training as of August 2025.

For AI scribes specifically, Supply Ontario procured tools through a vendor-of-record model and made them available to physicians, nurse practitioners, therapists, and other clinicians. The audit found that security, accuracy, and bias-related criteria carried limited weight in vendor evaluation. The process relied heavily on vendors' own claims rather than independent verification, with no requirement for live product demonstrations.

When Ontario Health and the provincial Ministry of Health tested 20 vendors, all exhibited at least one identified issue. More than half included incorrect medical information, including misidentified medical conditions and treatment recommendations. The audit also identified bias and hallucination risks.

Participants in the working group broadly agreed that AI scribes have moved from pilot to meaningful scale faster, and with more positive clinician reception, than most expected. But they acknowledged that public trust has not kept pace with adoption, and that oversight remains fragmented relative to international peers. The discussion centred on how Canada can safely accelerate broader healthcare AI commercialization.