Ontario's AI Medical Scribes Got Patient Names Wrong. That's Just the Start.

Walk into a North American clinic these days and you’ll notice something: the doctor is looking at you, not a laptop. The reason is a quiet AI scribe in the corner, listening in and drafting the chart in real time. Physicians have called it a burnout cure. An Ontario government audit just made that story a lot more complicated.

The exam-room AI you’ve probably already met

An AI medical scribe isn’t a glorified dictation app. It listens to the doctor-patient conversation, extracts clinical context — symptoms, diagnoses, prescriptions — and outputs a structured SOAP note (Subjective, Objective, Assessment, Plan). Tens of thousands of US and Canadian clinicians already use one, and Ontario went so far as to formally encourage adoption among family physicians to cut paperwork load.

The catch: that AI-generated note can land in a patient’s permanent medical record with little more than a tired doctor’s signature. When a clinician is running 15 minutes behind, “looks about right” becomes the de facto QA process — and an AI hallucination quietly becomes part of the chart forever.

What the audit actually found

The headline finding is brutal in its simplicity: the scribes are getting basic facts wrong.

Among the error types auditors flagged:

• Patient names swapped or misspelled
• Drug doses transcribed incorrectly (think 10mg recorded as 100mg)
• Symptoms appearing in notes that the patient never mentioned
• Diagnoses inserted that didn’t match what the doctor said
• Family history details bleeding in from other patients’ charts

The dosage errors are the scariest. In a domain where a single digit can kill someone, “the AI wrote it that way” is not a defense any malpractice court is going to accept.

Why this keeps happening

Strip away the marketing and you’re left with the core issue: large language models hallucinate. They predict plausible next tokens; they don’t verify facts. When the audio picks up a muffled “Tylenol 500,” the model can quietly “correct” it to 1000mg because that’s the more common dose in its training data. Statistical likelihood beats what the patient actually said.

Medicine makes this worse. Accents, cross-talk, mumbling, beeping monitors — speech recognition degrades fast under real clinical conditions. And rather than flagging uncertainty, generative models default to filling the gap with something confident-sounding. That’s the whole hallucination problem in one sentence.

The “human in the loop” that isn’t

Defenders of AI scribes lean hard on one argument: the doctor reviews everything before signing. On Hacker News and physician subreddits, the rebuttal is consistent — a family doc seeing 30 to 40 patients a day does not have time to line-edit every generated note. If they did, the productivity gain disappears, and the whole business case for the tool collapses.

So the operational reality drifts toward skim-and-sign. The audit is essentially a record of what slips through that gap.

Who eats the liability?

The real question the Ontario report forces is the boring one: who’s on the hook when an AI-authored note harms a patient? The doctor? The vendor? The province that recommended the tool?

Almost every scribe vendor’s terms of service push final responsibility onto the clinician. Doctors get the time savings; they also get the lawsuit. Medical malpractice insurers in the US and Canada are starting to ask pointed questions, and that conversation is going to get louder fast.

A familiar trap, new exam room

None of this means AI scribes are bad technology. Used carefully, they genuinely give doctors back attention to spend on patients instead of keyboards. The problem is running a healthcare system on the unstated assumption that “the AI wrote it, so it must be right.”

The Ontario audit is a reminder of an old pattern: when deployment outruns verification, the bill gets paid by the people with the least power in the room. Next time you see a small microphone next to your doctor, it might be worth asking — is that thing even getting my name right?