Ontario's healthcare auditors reviewed the AI note-taking tools that doctors use to document patient appointments - and found the tools are getting basic facts wrong, routinely.
The tools in question are called ambient scribes. They listen to a doctor-patient conversation in real time and automatically generate a clinical note - the written summary that becomes part of a patient's permanent medical record. The pitch to physicians has always been straightforward: instead of spending 30-plus minutes per day on documentation, focus on patients and let the AI write it up.
But that pitch assumes the AI gets things right. The Ontario audit found otherwise. The errors auditors flagged aren't edge cases or obscure technical failures - they're basic factual mistakes in notes that feed directly into diagnoses, prescriptions, and referrals.
The Real Problem Is the Workflow Around These Tools
Every software tool makes mistakes. What makes this finding dangerous is the workflow surrounding it. The entire value proposition of ambient scribes is that they reduce physician burden. That premise, in practice, discourages the careful line-by-line review that would catch errors before notes get filed. Doctors are time-pressed. When a tool promises to handle documentation, the natural response is to trust it - especially when charts are piling up.
Ambient scribes have spread fast through North American healthcare. Vendors have signed hospital-scale contracts over the past two years, often with minimal independent accuracy testing in live clinical environments. Adoption outpaced evaluation, and Ontario's auditors are now doing the evaluation after the fact.
A Pattern Showing Up Across High-Stakes AI Use
This is the same arc playing out in legal AI - tools adopted for their time-saving pitch, accuracy assumed rather than tested, and the problems surfacing when someone looks closely. Lawyers have filed AI-generated briefs containing fabricated case citations. Doctors are apparently filing AI-generated notes containing wrong facts. The category of problem is identical even if the domain is different.
For physicians using these tools today, the practical takeaway is unglamorous: every AI-generated note needs real review before it's filed. That probably cuts the time savings significantly, but it's the only way to catch errors before they become part of a permanent record.
For vendors, Ontario's audit is an early signal of where regulatory pressure is heading. If health ministries start requiring accuracy benchmarks or mandatory disclosure standards for clinical AI documentation tools, the current generation of products will face a harder bar than they passed to get into hospitals in the first place. The question is whether that pressure arrives before or after a documented patient harm.