Sick and wrong: Ontario auditors find doctors' AI note takers routinely blow basic facts

The robots are hallucinating in the exam room. And we're letting them write the permanent record.

Ontario's auditor general didn't mince words this week. AI-powered clinical note-taking tools — the ambient scribes that listen to patient visits and churn out SOAP notes — are getting basic facts wrong at rates that should embarrass a first-year med student. Wrong medications. Wrong dosages. Wrong diagnoses. In one case, an AI recorded a patient as having "resolved" a condition the doctor explicitly said was worsening. In another, it invented a family history of cancer that never came up.

This isn't a beta test. These tools are live in clinics across the province right now, feeding directly into electronic health records that follow patients for life.

The automation bias trap

Here's the sick part: doctors are signing off on these notes anyway.

The audit found clinicians reviewing AI-generated notes at speeds that suggest rubber-stamping, not verification. Median review time: 47 seconds. For a complex patient encounter that might have taken 20 minutes. That's not oversight. That's automation bias in a lab coat — the well-documented human tendency to trust machine output over our own judgment, especially when we're tired, rushed, and measured in RVUs.

The vendors knew this. They built for it. "Clinician-in-the-loop" is the marketing phrase, a fig leaf of accountability. But the loop is broken when the human in it has 15 patients before lunch and the AI serves up plausible-sounding prose that passes the sniff test at a glance.

We've been here before

Remember the Therac-25? The radiation therapy machine that gave massive overdoses because software replaced hardware safety interlocks? Remember the Boeing 737 MAX, where a single faulty sensor fed a system pilots weren't even trained to understand?

Every major automation disaster shares a DNA: efficiency pressure + opaque software + deprived human oversight = catastrophe. Healthcare IT has been replaying this loop for decades. CPOE alert fatigue. Copy-paste note bloat. Now ambient AI scribes — the same problem, supercharged by large language models that hallucinate with the confident tone of an attending physician.

The technology isn't the villain. The deployment model is.

The liability shell game

Ask the vendors who's liable when an AI error causes harm. They'll point to the "human in the loop" — the doctor who clicked "accept." Ask the hospital systems. They'll cite vendor contractslimiting liability to the subscription fee. Ask the regulators. In Ontario, the College of Physicians and Surgeons says doctors are responsible for their records. Full stop.

So the person with the least control over the algorithm bears 100% of the legal risk. The company that trained on scraped clinical notes, optimized for fluency over accuracy, and shipped without rigorous clinical validation? They cash the checks.

This is insane. And it's the standard playbook for AI in high-stakes domains: privatize the upside, socialize the downside.

What actual safety looks like

The auditor's recommendations are sensible but toothless: mandatory accuracy benchmarks. Human factors testing. Clear error-reporting channels. Vendor liability frameworks. All good. None binding.

Here's what binding looks like:

  • No AI note enters the legal health record without a documented, timed physician verification step — not a click, a dictated or typed attestation of specific clinical facts.
  • Vendors carry professional liability insurance proportional to deployment scale, with mandatory reporting of all clinically significant errors to a public registry.
  • Health Ontario funds independent, ongoing adversarial testing — red teams trying to break these systems in real clinical workflows, not vendor-supplied benchmarks on curated data.
  • Patients get notified when AI contributed to their record, and the right to flag discrepancies without fear of retribution.

Expensive? Yes. Slower? Initially. But the alternative is a silent corruption of the medical record — the substrate of every future clinical decision, every insurance determination, every malpractice case, every research study.

The trust tax

Patients already trust doctors less than they did a generation ago. They Google symptoms, they challenge guidelines, they record visits on phones. Now they'll learn — if the media does its job — that the note in their chart might be a confident fabrication by a model that has never seen a patient, never felt a pulse, never lost a patient.

Trust is the only currency medicine has. We're letting VC-backed startups spend it on our behalf.

The auditor called the findings "sick and wrong." That's not bureaucratic language. That's a warning. The question is whether Ontario's health leadership has the spine to treat it like one — or whether they'll issue guidelines, declare victory, and wait for the first preventable death to force their hand.

History suggests the latter. But the patients in those exam rooms didn't consent to be beta testers.