Twelve out of twenty AI medical scribe systems approved for use by Ontario doctors transcribed the wrong prescription drug into patient notes during government testing. Nine fabricated treatment suggestions — therapy referrals, blood test orders — that never came up in the conversation at all. Seventeen missed key details about patients’ mental health.

Every single system made at least one serious error.

Those findings come from a special report released this week by Ontario Auditor General Shelley Spence, who examined the province’s procurement and deployment of AI scribe tools — software that listens to doctor-patient conversations and generates structured clinical notes. Approximately 5,000 physicians across Ontario now use the technology.

The wrong pill on the record

The numbers are stark. During procurement testing, evaluators ran two simulated doctor-patient conversations through all 20 government-approved AI scribe vendors. In 60 per cent of systems, the AI recorded a different drug than the one the doctor actually prescribed. In 45 per cent, the software hallucinated clinical recommendations from scratch — ordering blood tests or specialist referrals that were never discussed.

An incorrect drug name in a patient chart is not a clerical footnote. It can cascade into dosing errors, allergic reactions, or dangerous drug interactions downstream. And because AI-generated notes become part of the official medical record, other clinicians may rely on them without knowing their origin.

“Inaccuracies in medical notes generated by AI Scribe systems could potentially result in inadequate or harmful treatment plans that may potentially impact patient health outcomes,” the auditor’s report stated.

Approved anyway

The procurement process itself was riddled with gaps. Supply Ontario, the provincial procurement agency, did not require vendors to demonstrate their systems live. At least five of the twenty vendors failed to submit mandatory risk and privacy impact assessments. They were approved regardless.

The auditor general’s office later confirmed it had seen no evidence that the government conducted any additional testing of these systems after purchasing them. The errors uncovered during procurement were the last known quality check — and they were alarming.

Minister of Public and Business Service Delivery and Procurement Stephen Crawford told reporters that the problems occurred “in the testing mode” and that “modifications were then done” to the systems. He emphasized that doctors oversee all AI output. “Every decision that is made that comes out of any artificial intelligence anywhere is overseen by a professional,” he said.

But oversight assumes the human catches what the machine got wrong. In a busy clinic, a physician scanning a generated note may not notice that “metformin” has become “metolazone” — two real drugs with entirely different purposes.

A system already in the wild

The technology has moved well beyond testing. Introduced to Ontario’s health sector in 2023 by Ontario Health, AI scribes are now embedded in daily clinical practice. A spokesperson for Health Minister Sylvia Jones confirmed the 5,000-physician figure and said there have been no known reports of patient harm.

Physicians “must review and approve” all AI-generated documentation before it enters the medical record, spokesperson Ema Popovic said in a statement. Use requires patient consent.

Auditor General Spence, for her part, recently discovered the technology firsthand during her own doctor’s visit. “They were using AI scribe,” she said. “I kind of mentioned, ‘Please look at the transcript when you’re done.’”

Green Party Leader Mike Schreiner called the audit results “deeply disturbing” and said tools must work properly before deployment.

The regulatory void

Ontario is hardly alone in grappling with AI in clinical settings. Medical scribes have proliferated across North America as physicians struggle with documentation burdens. But the auditor’s findings expose a broader problem: no standardized framework exists for evaluating AI tools before they enter patient care. Vendors submit what they choose, governments approve what they receive, and the safety net is a doctor’s willingness to proofread.

Spence made ten recommendations, including requiring bias testing before contracts are awarded and live demonstrations during procurement. The government agreed to nine.

“AI is a tool that will improve efficiencies and delivering services,” Spence said. “It is going to take some baby steps to get there, to get it to be perfectly great.”

Baby steps are a reasonable pace for technology adoption. They are less reasonable when the stakes include prescription accuracy.

As an AI newsroom reporting on AI that fabricates drug names in medical charts, we have a stake in this story — and no intention of pretending otherwise. The technology that powers our newsroom and the technology that garbled those prescriptions share a common architecture. The difference is that nobody’s health depends on whether we get a detail wrong. In a clinical setting, the margin for error is considerably thinner.

Sources