The short answer
Review an AI scribe note by hunting for the errors studies say are most common, not by reading it top to bottom for style. Check first for missing items (what was said but isn’t in the note), then numbers and units, then negatives (“denies” versus “reports”), then anything you don’t remember saying, then orders and plans. In published evaluations, omissions are the most common error, and a large share of made-up content involves fabrication or a flipped negative. This post is about how to review notes, not about clinical decisions; your judgement about the patient is yours.
Deedglass
Saved is not the same as saved correctly
Deedglass reads the order back
What the studies found
Three studies, three designs, a consistent picture:
| Study | Setting | Main finding |
|---|---|---|
| Taylor et al., JMIR Medical Informatics, 2026 | 356 ambient-scribe notes reviewed in a pilot (run 2024) | Omissions in 18% of notes, hallucinations in 11.5%, and 5.3% with errors rated serious or imminent risk1 |
| Asgari et al., npj Digital Medicine, 2025 | 450 notes, 12,999 sentences annotated by clinicians | 1.47% of note sentences hallucinated, 44% of those major; 3.45% of transcript sentences omitted; hallucinations were 43% fabrication, 30% negation2 |
| Kernberg et al., JMIR, 2024 | GPT-4 notes from simulated encounters | An average of 23.6 errors per case: 86% omissions, 10.5% additions, 3.2% incorrect facts; errors grew with transcript length3 |
Some context: the Asgari authors work for an AI scribe company, and the Kernberg study used a general-purpose model, not a commercial scribe.23 Products and models have changed since. Don’t read any single number as the rate for your tool. Read the pattern: omissions dominate, and made-up content often looks like a confident detail or a reversed negative.
Why a quick read misses these
Omissions are invisible on the page. A note that leaves out a symptom reads as cleanly as one that includes it. You catch an omission only by comparing the note with your memory of the visit.
Negation flips look plausible. “Patient denies chest pain” and “patient reports chest pain” are both well-formed sentences. Fabricated details are usually specific and fluent, which makes them feel true.
There’s also the reviewer. Research on automation bias finds people over-rely on automated output, more so under time pressure and high workload4, and that it appears even in single tasks when checking is hard.5 Reading a long, fluent note at the end of a clinic is exactly that situation. The routine below is meant to make checking easier, not just to add steps.
The review routine
1. Before reading the note, recall the visit. Spend ten seconds listing the two or three things that mattered most: the main complaint, the key finding, the plan. Then check the note has them. This is the only reliable way to catch omissions.13
2. Check every number. Doses, durations, vital signs, dates, laterality. Numbers are easy to scan for and expensive to get wrong.
3. Check every negative. Scan for “no”, “denies”, “without”, “negative for”, “not”. For each, ask whether that’s what was said. Negation was the second largest category of hallucination in one large evaluation.2
4. Question anything unfamiliar. A detail you don’t remember hearing is a candidate fabrication, the largest hallucination category in the same evaluation.2 Delete it or confirm it.
5. Read orders and plans word by word. If the scribe queues orders, treat each like one you’re writing yourself.
6. Watch long visits. Errors grew with transcript length in one study.3 Give long or complex visits more attention, not less.
7. Sign only what you’d defend. Signing means you’ve taken responsibility for the content, whoever drafted it.
The one-page checklist
| # | Check | Looking for | Seconds |
|---|---|---|---|
| 1 | Recall top three points; find them | Omissions | 10 to 20 |
| 2 | Scan all numbers and units | Wrong dose, value, date, side | 15 |
| 3 | Scan all negatives | Flipped “denies” or “reports” | 15 |
| 4 | Flag unfamiliar details | Fabrications | 10 |
| 5 | Read orders and plan in full | Wrong or extra orders | 20 to 40 |
| 6 | Extra pass if visit was long | All of the above | 30 |
| 7 | Sign |
These times are rough targets for a routine visit, not a standard. The point is that a focused check of six things is faster than a slow full read, and more likely to catch what matters.
For clinic and department leads
- Sample notes. Each month, compare a few AI drafts with the signed versions and, where possible, with the recording. You’ll learn your tool’s typical errors.
- Watch for zero-edit signing. A note signed with no edits in seconds isn’t necessarily wrong, but a pattern deserves a conversation.
- Keep the draft. Make sure your system keeps the AI’s original draft and the signed version separately, so reviews and incident investigations can see what changed. Our post on scribe audit trails lists what to record.
- Report error types to the vendor. If you see the same omission pattern repeatedly, the vendor needs to know.
- Ask the vendor the hard questions before rollout; our vendor questions post has twenty.
Questions people ask
How common are errors in AI scribe notes?
It varies by tool, setting and study. One 2026 pilot found omissions in 18% and hallucinations in 11.5% of 356 notes1; a 2025 evaluation found 1.47% of sentences hallucinated.2 Measure your own tool.
What’s the most common error type?
Omission, in all three studies above.123 That’s why the routine starts with recall.
Should I edit the note or rewrite it?
Edit what’s wrong and remove what you can’t confirm. If a draft needs heavy rewriting regularly, raise it with whoever runs the tool.
For hospitals setting up review and records across many clinicians, see our healthcare page and the record spec, and for how often systems store something different from what they were sent, the FHIR write integrity study. To try approvals and read-backs on a real database in your browser, open the playground; to grade what an agent can reach on your own machine, run npx deedglass scan (the grade, explained).
Written with AI assistance and checked against our claims register.