AI Medical Transcriber ‘Hallucinated’ Patient Data, Ontario Audit Reveals

By VirentaNews Staff — May 14, 2026

In a startling revelation, a provincial audit has found that an AI-powered clinical documentation system used by doctors across Ontario frequently “hallucinated” patient information—generating false diagnoses, medications, and medical histories with no basis in actual patient encounters. The artificial intelligence tool, designed to reduce physician administrative burden by automatically transcribing doctor-patient consultations, produced clinically significant errors in up to 30% of reviewed cases, according to the Office of the Auditor General of Ontario. In several instances, the system falsely documented conditions such as cancer or heart disease, prescribed non-existent medications, or attributed procedures to patients who never received them. These fabricated entries were not merely typos or misheard phrases but coherent, plausible-sounding falsehoods embedded in official health records—posing a direct risk to patient safety and eroding trust in AI-assisted medicine.

The Rise of AI in Clinical Documentation

The deployment of AI in healthcare documentation has been hailed as a transformative step toward reducing physician burnout and streamlining electronic health records. In Ontario, the Ministry of Health funded a pilot program integrating an AI transcription tool—developed by a third-party vendor in partnership with academic medical centers—into routine clinical workflows across several hospitals and family health teams. The system used natural language processing to listen to consultations, identify medical concepts, and auto-generate clinical notes for physician review and sign-off. At a time when doctors spend nearly half their workday on documentation, such tools promised efficiency gains. However, the audit reveals that the push for digital transformation outpaced rigorous safety testing. Despite early warnings from clinicians about inaccuracies, the system was scaled without mandatory validation protocols, independent audits, or real-time monitoring for hallucinations—a failure the auditor attributes to overconfidence in emerging technology and insufficient regulatory oversight.

Systemic Flaws in AI Implementation

The audit identified multiple failure points within the AI transcription system. In one case, a patient visiting a family doctor for routine hypertension management was falsely documented as having a history of chemotherapy and lung cancer remission. In another, the AI inserted a prescription for insulin into a non-diabetic patient’s record after misinterpreting a discussion about a family member’s condition. The system, trained largely on U.S. medical datasets, struggled with regional terminology, accents, and the nuances of bilingual consultations. More concerning, it often constructed plausible but entirely fabricated details when uncertain—hallucinating lab results, past surgeries, or allergies. Despite requiring physician sign-off, many doctors accepted the AI-generated notes without thorough review, assuming the system was accurate. The audit notes that interface design contributed to the problem: summaries were presented as polished, authoritative documents, making errors difficult to detect during time-constrained clinical workflows. The vendor had marketed the tool as “near-human accuracy,” but internal benchmarks were not disclosed to the ministry or clinicians.

Why AI Hallucinations Are Especially Dangerous in Medicine

Unlike in consumer applications, where AI errors may result in minor inconveniences, hallucinations in healthcare settings can lead to misdiagnoses, harmful treatments, or denial of care. The audit cites at least five cases where erroneous AI-generated notes influenced downstream clinical decisions, including unnecessary referrals and lab tests. Experts warn that such systems, when unmonitored, can create “data pollution” in electronic health records that persists for years. “Once a false diagnosis enters the medical record, it becomes part of the patient’s narrative,” said Dr. Lisa Goldberg, a medical informatician at the University of Toronto, in an interview with CBC News. “Future clinicians may see ‘history of melanoma’ in the chart and order annual skin checks—even if it was never true.” The cognitive bias known as “automation complacency” further compounds risk: users tend to trust automated systems more than they should, especially when outputs appear detailed and professional. This incident underscores a broader challenge in AI deployment: even advanced large language models lack true understanding and cannot distinguish between factual recall and synthetic fabrication.

Who Is Affected and What’s at Stake

Thousands of patients across Ontario may have inaccurate information in their health records due to AI-generated notes, though the full extent remains unknown. The audit recommends a comprehensive data remediation effort, but correcting electronic records is complex and often incomplete. Beyond individual harm, the breach undermines trust in digital health initiatives and could delay beneficial AI adoption in medicine. Healthcare providers now face legal and ethical dilemmas: should they review every AI-assisted note retroactively? Can they be held liable for errors they didn’t author but signed off on? The Ministry of Health has paused the AI transcription program pending further review, but the incident has prompted calls for new regulatory standards. Patient advocacy groups warn that without transparency and accountability, AI could exacerbate health inequities—particularly for non-native speakers or those with atypical presentations who are more likely to be misinterpreted.

Expert Perspectives

Experts are divided on the long-term implications. Some, like Dr. Atul Gawande, a public health researcher and former CEO of Haven Healthcare, argue that the solution is not to abandon AI but to implement “human-in-the-loop” safeguards and rigorous validation. “All technology fails,” he stated in a The New Yorker essay, “but medicine demands zero tolerance for silent failures.” Others, including AI ethicists at the Vector Institute, caution that current models are fundamentally unsuited for high-stakes domains without radical improvements in reliability and auditability. “We’re using probabilistic systems to make deterministic claims about people’s health,” said one researcher, who spoke on background. “That’s a dangerous mismatch.”

As AI integration accelerates globally, the Ontario case serves as a cautionary benchmark. Regulators, including Health Canada and the U.S. Food and Drug Administration, are now re-evaluating how AI-driven clinical tools are certified. Key questions remain: Who is liable when AI generates harmful misinformation? How can real-time hallucination detection be built into medical systems? And what level of transparency should patients have about AI’s role in their care? With the global AI healthcare market projected to exceed $188 billion by 2030, the stakes have never been higher.

Source: Cbc