A provincial audit has revealed that the AI transcription tools Ontario approved for its healthcare system are fundamentally broken. The Office of the Auditor General of Ontario evaluated 20 vendors in the government's AI Scribe program, which is designed to help physicians document patient encounters automatically. What auditors found should concern anyone who's ever sat in a doctor's waiting room: these systems hallucinate content, misidentify medications, and routinely miss critical mental health information that patients explicitly discussed during appointments.

The Scope of the Problem

Nine out of twenty AI systems evaluated "fabricated information and made suggestions to patients' treatment plans" that never appeared in the original doctor-patient recordings. Evaluators flagged specific dangerous errors: notes indicating no masses were found when nothing about masses was ever mentioned, and patient anxiety noted despite zero discussion of mental state during consultations. Twelve of the twenty systems inserted incorrect drug information into patient records, while seventeen systems missed key details about patients' mental health issues that were clearly discussed in the recordings. Six systems "missed the patients' mental health issues fully or partially."

The Evaluation Failures Go Deeper Than Bad AI

Here's where it gets truly absurd. According to the report, Ontario's procurement scoring heavily weighted criteria unrelated to clinical accuracy. A vendor's domestic presence in Ontario contributed 30 percent of their total evaluation score. Meanwhile, the accuracy of medical notes contributed just 4 percent. Bias controls accounted for only 2 percent, threat and privacy assessments also weighed in at 2 percent, and SOC 2 Type 2 compliance added another 4 percentage points. In other words, criteria directly tied to patient safety made up a single-digit percentage of how these systems got approved for use across the province's healthcare system.

No Safety Net for Physicians

OntarioMD, the group involved in AI Scribe procurement, has recommended that doctors manually review their AI-generated notes for accuracy. But here's the kicker: none of the twenty approved systems include mandatory attestation features that would require physicians to confirm they've reviewed and verified the documentation. More than 5,000 Ontario physicians are currently using these tools, and the Ministry of Health told the CBC it has received no reports of patient harms associated with the technology so far.

Key Takeaways

  • Nine of twenty AI Scribe vendors fabricated treatment suggestions not discussed in appointments
  • Twelve systems inserted incorrect medication information into patient records
  • Medical note accuracy contributed only 4% to procurement scoring while Ontario presence contributed 30%
  • Over 5,000 physicians using approved tools with no mandatory review requirements

The Bottom Line

This is exactly the kind of bureaucratic failure that makes security researchers bang their heads against walls. You can't hand doctors broken transcription tools, weight vendor location over accuracy during approval, skip mandatory attestation, and then act surprised when the systems routinely hallucinate clinical details. Patient safety isn't a feature to be scored at 4 percent.