In June 2026, the Canadian government released its long-awaited National AI Strategy, which identifies health care as its “first mission.” The strategy contends that artificial intelligence (AI), if “deployed well,” can “expand access to primary care, reduce ER wait times, prevent avoidable visits through better upstream care, and lighten the administrative burden on physicians” (p. 29). As examples, the strategy highlights two “AI scribes” operating in Alberta — OKAKI’s CliniQuill and Jenkins — the former facilitating community health services in Indigenous communities and the latter used in emergency departments across the province.
AI scribes are tools, typically based on large language models that record and transcribe conversations between medical practitioners and patients. The tools generate structured, regulatory-compliant notes of the discussion that include diagnoses, treatment plans such as blood tests and prescriptions that become part of the patient’s medical record. Doctors, nurses and even some paramedics in Canada are using AI scribes.
While the goal to expand Canadians’ access to health care is laudable, some health officials are sounding alarm at health-care practitioners using AI scribes. The Canadian Medical Association (CMA) warned the federal government in 2024 that “the adoption of AI applications in health care” is “outpacing dedicated regulation” (p. 1). Similarly, officials in Australia’s federal health department stated that AI scribes have “little oversight” with serious “implications for patient safety, clinical accountability, and the integrity of data.” University of Melbourne law associate professor Megan Prictor calls for regulation of the “wild west” of AI medical scribes.
Canada is witnessing serious problems with AI scribes’ data accuracy. In a May 2026 report, the Ontario auditor general evaluated 20 AI scribe tools that were approved by Supply Ontario, the government procurement department. The report found inaccuracies in AI transcriptions including errors, incomplete information and AI hallucinations. For example, nine of the 20 approved tools (45 percent) fabricated information such as ordering blood tests which was not present in recordings, while notes generated by 12 of the 20 tools (60 percent) captured a drug different than what was prescribed by the doctor (p. 23). Notes generated by most tools (17 of the 20, 85 percent) missed key details about the patients’ mental health.
Peer-reviewed studies confirm consistent problems with AI scribes, such as in a 2025 article in the journal JMIR Human Factors that tested six AI scribes in Toronto. Researchers found that while the tested tools produced “good to excellent quality medical notes, none were consistently error-free.” The researchers noted that transcription accuracy can be complicated when there are interruptions, loud background noises and multiple speakers. A 2024 report commissioned by the Ontario Ministry of Health had similar findings, with primary care doctors reporting AI scribes reduced their paperwork burdens but had limitations in transcribing multilingual conversations, speakers with accents or enunciation difficulties, or complex multi-issue appointments. These are not uncommon occurrences in health-care settings, especially in emergency departments.
Inaccuracies or omissions in AI-generated medical notes could potentially cause serious problems if patients receive inappropriate, inadequate or harmful treatment plans.
Fixing the Oversight Gap
Equally problematic are gaps in governance. The Ontario auditor general’s report noted that the 20 AI scribes that Supply Ontario pre-qualified for use in Ontario were not comprehensively evaluated yet were still approved. Eleven of those 20 approved vendors did not submit required third-party auditor reports or certification from the International Organization for Standardization. Instead, Supply Ontario relied upon vendors’ self-reported data. In addition, Supply Ontario did not require vendors to conduct live demonstrations of the tools to test efficacy, but allowed vendors to submit simulated recordings and attest that the recordings were unedited. Five approved vendors were also allowed to simply claim that they met privacy and security requirements, as they had not submitted the required threat risk assessments and privacy impact assessments.
Shockingly, one Supply Ontario-approved scribe even seemingly fabricated testimonials from non-existent doctors, according to an investigation by TVO and the Investigative Journalism Foundation. Supply Ontario’s reliance upon vendors’ self-reported data raises serious questions about the integrity of its approval processes.
Health practitioners have a professional responsibility for ensuring that medical notes, including AI-transcribed notes, are accurate and complete before adding them to patients’ records. This critical role of doctors reviewing all AI scribe notes was underlined by OntarioMD — a subsidiary of the Ontario Medical Association, funded by the Ontario Ministry of Health and Long-Term Care — in its response to the auditor general’s report.
But physician review of AI scribes’ outputs does not address some AI scribes’ privacy and security problems identified by the Ontario auditor general, nor Supply Ontario’s reliance upon vendor-supplied data.
To address broader governance problems related to AI-powered tools in health care, the CMA issued recommendations in October 2025 for the federal government. These recommendations include mandating developers to disclose how health systems are “trained, validated and monitored” during development and after implementation (p. 4). The CMA also recommends that Canada adopts standardized pathways for the validation of AI tools through real-world testing and post-market surveillance to ensure the safety and efficacy of the software, and it stresses the importance of independent evaluation by qualified third parties, not vendor self-reporting.
The CMA’s recommendations, which accord with the Ontario auditor general’s report, are a useful step to address the problem of regulators’ overreliance on vendors’ data and claims about their products. Supply Ontario responded that it will follow the report’s recommendations to improve its procurement processes.
Supply Ontario’s commitment is a welcome step that will help ensure approved vendors have independently verified security, privacy and quality-control standards. Its processes, however, currently apply to a small number of vendors. Many AI scribes are not assessed by Supply Ontario or other regulatory bodies. As a result, the burden for determining the scribes’ security, privacy, data governance and accuracy falls to individual health-care practitioners and clinics, highlighting the importance of regulators or peak bodies establishing lists of vetted vendors.
In addition to serious governance concerns, AI scribes raise challenges regarding liability for errors when an AI scribe is used. The CMA, for example, warns doctors to understand how the AI scribe assigns liability, such as if a company uses blanket clauses that shift all risk to the doctor or whether the vendor accepts responsibility for privacy or security failures. Here, Canada can look to Australia for concrete proposals, as Australia has held a national expert consultation on AI in health care in 2024, in which clinician liability emerged as a core concern.
The Royal Australasian College of Physicians (RACP), which represents doctors across Australia and Aotearoa New Zealand, warns that doctors are in a “legal ‘black hole’” in relation to AI tools and doctors’ liability. The RACP recommends refinements to “existing governance and regulatory frameworks at local, state and national levels” in its 2026 position statement, so that doctors do not carry primary responsibility for patient safety when AI tools are in use (p. 4).
In its 2025 report on its public consultation on AI and health care, the Australian government has acknowledged its “current regulatory system is not fit for purpose” and is considering a range of regulatory measures, including “mandatory AI guardrails” that could incorporate data governance measures and supply chain transparency (p. 8). This is the kind of whole-of-government response that Canada needs to consider to ensure effective regulatory frameworks for AI in health care. There is also a critical need in Canada to coordinate amongst all levels of government and, importantly, to consult publicly with health-care bodies, practitioners and academic experts for regulation that is appropriate for high-risk settings such as health care.
That health care is one of Canada’s “first missions” using AI is welcome news to our over-burdened and underfunded public health systems. Before we celebrate, and to reap the rewards of this burgeoning market, we need oversight that matches the ambition.
