Ontario’s audit of 20 approved AI Scribe vendors found 60% mixed up prescribed drugs and 45% fabricated clinical information never discussed in patient recordings.
Key Takeaways
12 of 20 AI Scribe systems inserted incorrect drug information; 9 of 20 fabricated treatment suggestions not present in simulated recordings.
17 of 20 systems missed key mental health details; 6 missed mental health issues fully or partially.
Procurement scoring weighted Ontario domestic presence at 30% of total score; medical note accuracy counted only 4%.
No mandatory attestation feature exists in any approved AI Scribe system despite OntarioMD recommending manual review.
Over 5,000 Ontario physicians are currently using these systems; the Ministry reports no known patient harms so far.
Hacker News Comment Review
Commenters with firsthand experience confirmed hallucination in live clinical and workplace settings, with fabricated diagnoses and invented vendor commitments appearing in real notes.
The procurement scoring breakdown drew the most technical criticism: weighting domestic presence 7.5x higher than accuracy is seen as the structural root cause, not just model quality.
Several commenters raised the baseline question of human error rates in medical records, noting no comparative data was provided, which weakens the implied severity framing.
Notable Comments
@Groxx: Personally received an AI summary diagnosing him with osteoporosis and hip pain after a Runner’s Knee visit – none of it was said. “CHECK YOUR TRANSCRIPTS.”
@Hobadee: Notes timestamped to recording segments enable spot-checking; argues provenance linking is the minimum viable safety feature for clinical AI scribes.
