How AI summaries of medical records fail: 10 failure types
Answer in brief: AI record summaries fail in 10 ways, which fall into 4 groups: facts left out, meaning changed (negatives, hedges, who said it), facts put in the wrong time or place (dates, sides, doses), and defects in the file itself (copied-forward text, another patient's pages). Each type has a cue you can search for, so you don't have to reread the file to find it.
Limits: We can't vouch for any tool's error rate on your files: the studies we cite measured other tasks and other models, and vendor features are as each vendor's site stated on the capture date.
Every example below is synthetic, no real patient. Each pair shows the source wording on the left and a plausible wrong summary line on the right. The examples come from the synthetic test file, so you can see them in context.
Facts left out: omissions
1. Omission
The summary is missing a fact that matters to the question: a prior complaint in the same body part, an abnormal lab, a consult that changed the plan. Omission is the most common error in every study on the evidence page, and the hardest to see, because every line that is there can be right.
Why it happens: the model judges the fact unimportant, or the fact sits in a part of a long input the model uses poorly. Research on general language models found they use the start and end of a long input better than the middle.
How to catch it: run a reverse spot check: pick source pages at random and confirm their key facts appear in the summary. Then search the source for dates before the injury or event and read those pages yourself.
2. Invented fact
The summary states something the file doesn't contain. It is rarer than omission in the studies, and the easiest to catch when every line carries a citation, because the cited page won't say it. The dangerous version is the plausible invention: the result of a test that was ordered but never resulted.
How to catch it: for every result in the summary, find the result report, not the order. Search the source for "ordered", "scheduled" and "pending", and check that each has a matching report or a missing-record flag.
Meaning changed: negatives, hedges and who said it
3. Lost negative
A negated finding comes out positive. In the largest annotated study, 30% of hallucinated sentences were negations. It happens because the model predicts likely words, and in a shoulder injury file "numbness" is likely whether or not the patient denied it.
How to catch it: search the source for denies, no, negative, without, absent and "not seen", and confirm each negative on a relied-on finding survived. The dates and negatives page has the full cue list.
4. Hedge dropped
Clinical writing marks uncertainty: rule out, possible, probable, cannot exclude, suspicious for, likely. A summary that drops the hedge turns a differential into a diagnosis. None of the studies on the evidence page counts it as its own category, but it is a close cousin of the negation error and it matters in any review that turns on when a diagnosis was made.
How to catch it: search the source for r/o, rule out, possible, probable, versus, vs, cannot exclude, concern for, suspicious for, likely and questionable. For each diagnosis in the summary, find the first page where it appears without a hedge.
5. Attribution error
The summary changes who said it. A patient's account becomes a clinical finding, or a patient's report of what another doctor said becomes that doctor's recommendation. In medical-legal work this matters as much as the fact itself, because history and findings carry different weight.
How to catch it: search the source for pt states, reports, per patient, according to, told, and "per" followed by a name. Check that the summary keeps the speaker. Then look for the other doctor's own note: here, the orthopedic addendum on p. 12 says there is no surgical indication.
Wrong time or place: dates, sides and doses
6. Wrong date
A single page can carry 5 or more dates: the fax header, the print date, the date of service, the dictation date, the signature date, and for labs the collection and result dates. A tool that picks the wrong one moves a visit, reorders events or invents a gap. The date-type field guide lists which date to use for which question.
7. Wrong side or level
Right becomes left, a finger becomes the wrong digit, L4-5 becomes L5-S1. These errors come from OCR on poor scans, from a stray line in an otherwise consistent note, or from a co-mingled page. They change the case.
How to catch it: search the summary for every left, right and bilateral on a relied-on finding and compare each to its page. Check the side against the injury report and the first visit.
8. Dose or unit error
Doses fail on combination products, on handwriting and on OCR. "5/325" is 5 mg of hydrocodone with 325 mg of acetaminophen; a summary that writes "hydrocodone 325 mg" is wrong by a factor of 65. In a 2026 study of AI summaries of 30 synthetic discharge letters, the medication section scored lowest and had the most hallucinations.
How to catch it: check every dose the case depends on against the page image, not the extracted text. Look at the units, the frequency and whether the product is a combination.
Defects in the file: copy-forward and wrong patient
9. Copy-forward echo
EHR notes repeat earlier text. In a study of 104 million notes at 1 health system, 50.1% of all note text was duplicated from earlier text about the same patient. A summary that trusts the latest note can report a work status, a plan or an exam that was copied from a visit weeks earlier and has since changed.
How to catch it: look for "unchanged", "no change", "as above" and paragraphs that repeat word for word across visits. When a note and a form disagree, the summary should flag the conflict, not pick 1.
10. Wrong patient
Large productions sometimes include another patient's pages: a similar name, the same imaging center, the same date of service. A summary that reads them as your patient's adds a diagnosis or a surgery the patient never had. A perfect language model makes this error too, because the page really does say it.
How to catch it: check name, date of birth and record number on the first page of every new source. Transposed digits and near-identical names are the pattern.
Which failure types matter most for your review
All 10 can occur in any file, but the ones that change an opinion differ by review type. Use this to decide where your check time goes.
| Review | Check first | Why |
|---|---|---|
| IME or QME | Omission (prior history), attribution, copy-forward work status | Apportionment and causation turn on prior history and on what the patient reported vs what was found |
| Standard-of-care review | Wrong date, hedge dropped, lost negative | The question is what was known when; a dropped "rule out" moves the date a diagnosis was made |
| Peer or utilization review | Invented fact (ordered vs resulted), dose and unit | Medical necessity rests on results and treatment actually given |
| Transfer or consult prep | Dose and unit, lost negative, omission (allergies) | Care decisions follow directly from the summary |
| Any large production | Wrong patient, duplicates | Defects in the file pass through any tool unchanged |
This table is our judgment as of 2026-11-12, based on what each review type has to decide. It is not drawn from a study.
Questions about AI summary failure types
What is the difference between an omission and a hallucination?
A hallucination puts something in that the source doesn't support. An omission leaves out something the source does contain. Studies count them separately because they need different checks: a hallucination fails when you open its citation, while an omission has no citation to open.
Why do AI summaries flip negatives?
Language models predict likely text. In a file about shoulder pain, "numbness" is a likely word, so a "denies numbness" can lose its negation in a long summary. 30% of hallucinated sentences were negations in the 2025 npj Digital Medicine study.
Can a better model fix wrong-patient pages and copy-forward?
No. Those are defects in the file. A model that summarizes faithfully will repeat them. They need checks on the file itself: identifiers on every source, and conflicts between notes and forms flagged.