Your Doctor’s AI Notetaker is Hallucinating: Ontario Audit Reveals 15% Error Rate

An Ontario healthcare audit just confirmed what many of us feared: your doctor’s AI medical notetaker audit reveals these tools are literally making things up. The report, released May 15, 2026, found that transcription errors and hallucinations occurred in roughly 15% of recorded patient visits. This isn’t just a minor typo in a Slack message; it’s a permanent error in your medical history. If your doctor uses tools like Nuance DAX or Nabla, you need to pay attention to your charts.

The Hallucination Problem in Clinical Settings

The audit reviewed over 2,000 AI-generated summaries across Ontario clinics. It found that LLMs like GPT-4 and Gemini 2.0, which power many of these apps, occasionally fill in the gaps when audio is muffled or medical terminology is complex. In one documented case, the AI recorded a patient as having chronic heart failure when they actually discussed seasonal allergies. These tools cost clinics between $100 and $300 per month per clinician. Microsoft’s Nuance DAX Copilot is a market leader here, but even with its high monthly price tag, it isn’t perfect. The audit highlights that human-in-the-loop verification is failing because doctors are too busy to proofread 500-word summaries. I’ve seen this in tech for years: we ship first and fix later, but fixing later doesn’t work in a surgery prep room.

Why LLMs fail in clinical settings

LLMs are probabilistic, not deterministic. They predict the next likely word, which is great for a blog post but dangerous for a prescription dosage. When the AI hears 50mg, it might hallucinate 150mg based on common patterns in its training data, creating a life-threatening error in a patient’s digital file.

The High Cost of Administrative Efficiency

Doctors are burnt out. The average GP spends 2 hours on paperwork for every 1 hour of patient care. Naturally, they turned to AI. Nabla, which offers a free tier and a $119 per month Pro version, has seen massive adoption across North America. But the Ontario audit shows that speed comes at a price. The report notes that 1 in 10 AI-generated notes contained clinically significant errors. We are talking about incorrect medications or missed allergy warnings. While Microsoft claims DAX saves 7 minutes per patient, that time saving is negated if the doctor has to spend 10 minutes auditing the AI’s work. I think we are rushing to replace scribes with software that isn’t ready for the high-stakes environment of a hospital.

The $30,000 administrative burden

Clinics are spending upwards of $30,000 annually on AI subscriptions and specialized hardware, like high-end directional microphones, to improve accuracy. Despite this investment, the error rate remains stubbornly high in noisy environments where multiple people are speaking during a consultation.

Privacy Risks and Data Sovereignty Concerns

It’s not just about accuracy; it’s about where that data goes. The audit found that 22% of clinics didn’t clearly disclose where AI data was stored. Many of these startups use APIs from OpenAI or Anthropic. While Enterprise agreements usually promise data won’t be used for training, the audit found shadow IT where doctors used consumer-grade ChatGPT Plus (the $20 per month version) to summarize notes. That is a massive HIPAA and PIPEDA violation. If your data is being fed into GPT-4o to improve the model, your private health info is effectively out in the wild. I always check if a service uses Zero Retention APIs, but most patients don’t even know their voice is being recorded in the first place.

Risks of consumer-grade AI in clinics

Using a standard iPhone 16 Pro to record a session via a basic AI app is a disaster. Professional tools like Suki or Abridge use encrypted silos, but the audit shows many clinics take shortcuts to save on subscription costs, putting patient privacy at risk.

What This Means for Your Medical Records

You have the right to see your notes. Under current laws in the US and Canada, you can request your medical records at any time. I recommend doing this after every major appointment. Look for the AI-generated disclaimer at the bottom of the note. If it’s there, read every line. The Ontario audit suggested that patients should be given a copy of the AI summary before leaving the office. This would catch the 15% error rate before it hits the insurance billing cycle. If you see an error, demand a correction immediately. Don’t let a hallucinated diagnosis sit on your record for years; it will affect your insurance premiums and future care. It is much harder to delete a false diagnosis than it is to prevent it.

How to request your digital health records

Most modern providers use portals like MyChart. Log in, check the Notes section, and look for discrepancies between what you actually said and what the AI wrote. If your doctor uses Epic or Cerner systems, these notes are usually available within 24 hours.

⭐ Pro Tips

• Ask your doctor specifically: Is an AI transcribing this, and which model are you using? If they say ChatGPT, ask them to use a human scribe or traditional notes.
• Always use a patient portal like MyChart to review notes within 24 hours of your visit to catch hallucinations early.
• If you find a mistake, send a secure message through the portal immediately so there is a timestamped record of your correction for insurance purposes.

Frequently Asked Questions

Final Thoughts

AI is great for writing emails or generating code, but it’s clearly not ready to be your unofficial medical scribe without heavy supervision. The Ontario audit proves that the 15% error rate is a systemic risk that can’t be ignored. Don’t be a passive patient. Treat your medical records like your credit report—check them often and dispute errors immediately. AI is a tool, not a doctor, and it still needs a human to check its work.