Before you sign the next note your AI scribe drafted, know this: the single largest published study of the practice found key clinical details -- a medication, an allergy, a finding the patient actually reported -- missing or wrong in roughly 26% of cases, and a separate 2026 audit of real deployed notes found that headline accuracy figures move almost entirely with how strictly reviewers are told to look, not with which product is being graded. Neither fact is a reason to distrust AI scribes like Heidi Health's or Abridge's wholesale -- clinicians are adopting them fast because the time savings are real. It's a reason to know exactly what to check before your signature makes the note the official record, because that signature is what the responsibility actually rests on.
What 'AI scribe accuracy' numbers actually measure
Three independent studies have measured AI-scribe error rates in the past year, and they don't agree with each other -- not because the tools disagree, but because the studies measured different things. A comparison of four commercial scribes against two simulated internal-medicine and surgical encounters, published in Studies in Health Technology and Informatics, found 71% of all errors were omissions -- information that was said but never made it into the note -- against 19.4% additions and 6.5% incorrect facts. Mayo Clinic Proceedings: Digital Health ran a larger test in October 2025: five ambient scribe platforms against 14 simulated ambulatory encounters, finding 26.3% of key clinical elements per case omitted or inaccurately captured, with omissions again the majority error type at roughly 76.3%, and an average of 3.0 errors per case carrying potential for moderate-to-severe patient harm.
The third study is the one that should change how you read the other two. A September 2026 preprint audited 565 real notes from 142 actual consultations across three deployed commercial scribes -- not simulations -- using an adversarial two-model review panel, and found 31.3% of notes carried at least one independently verified failure, concentrated in allergy and medication information, invented patient identity, and physical-exam language written into notes from telephone-only visits. But the authors' own omission share of their verified errors was just 23.1% -- far below the 54-86% range other published audits report -- and when they held the scribe, the evidence, and the settings fixed and changed only the instruction given to reviewers, the share of candidate errors that counted as verified moved from 9.3% to 79.0% on the exact same data. The instrument used to measure an AI scribe's accuracy determines the number it reports almost as much as the scribe itself does.
What each scribe-accuracy number actually counts
- 71% · share of errors that are omissions
- 4 commercial scribes, 2 simulated encounters (IOS Press study)
Includes: Errors found comparing AI notes to a reference transcript on scripted, simulated visits.
Excludes: Real deployed notes, and any measure of which omissions were clinically dangerous. - 26.3% · of key elements omitted or inaccurate per case
- 5 platforms, 14 simulated ambulatory encounters (Mayo Clinic Proceedings)
Includes: Omissions and inaccuracies together, plus a separate finding that ~76.3% of all errors were omissions.
Excludes: Real patient encounters -- every case here was simulated. - 31.3% · of notes with a verified failure
- 3 deployed scribes, 565 real notes, 142 real consultations (arXiv audit)
Includes: Real-world deployed notes, adversarially verified by two separate AI reviewers plus human spot-checks.
Excludes: A severity ranking -- and their own omission share of errors was 23.1%, not the 54-86% other audits found.
The five-minute check before you sign
None of the three studies above will tell you whether the specific note in front of you right now is one of the ones with an error. What they do tell you, consistently, is where to look first.
Check an AI-drafted clinical note before you sign it
- Medication and allergy information is the single category every study here flags as highest-harm -- and the category Heidi Health's own published safety guidance separately names as a known failure point for speech recognition.
- Omission is the dominant error type across all three studies -- as high as 76.3% of errors in one -- and a missing detail reads as a clean, confident note, exactly like a wrong one does.
- The largest real-notes audit to date found physical-exam language written into notes from telephone-only consultations specifically -- a fabrication a remote visit makes structurally impossible to have happened.
- The same audit found invented patient-identity details among its most common verified-failure categories -- a scribe filling an inaudible stretch of audio with a plausible-sounding guess rather than leaving a gap.
- Research cited in the Mayo Clinic Proceedings study found clinicians reviewing AI-drafted text on screen frequently failed to catch clinically relevant errors -- a skim is a weaker check than it feels like.
Running that check closes the specific gaps three separate studies actually found. The ways it quietly gets skipped anyway are mostly the same few, every time.
Four ways this check gets skipped without anyone noticing
None of this is an argument against AI scribes -- the documentation-time savings behind their fast adoption are real, and independent clinician guidance on Heidi Health's own product is explicit that the vendors aren't disputing any of this either. It's an argument for treating every AI-drafted note the way you'd treat a draft written by a bright new hire who wasn't actually in the room: useful, fast, and never the final word until you've checked it against what you know happened. Accuracy is one half of the AI-notetaker question; whether everyone in the room agreed to being recorded in the first place is the other, and it's the half with an actual lawsuit attached.
- A 2025 study found AI scribes omitted or garbled roughly one in four key clinical details.
- Omission, not garbled text, is the dominant error type in every published scribe-accuracy study.
- A 2026 audit found published failure rates swing from 9% to 79% on identical notes by instrument.
- Check medications, allergies, and telehealth exam claims first -- they carry the most real harm.
- Caveat: a vendor's own accuracy percentage is not a substitute for reviewing the note yourself.