Glossary: terms for AI summary errors and checks

By Ahmed Jemaa, Co-Founder and CEO of Medrecords AI. Updated 2026-11-12.

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.

Error types

Hallucination

A statement in an AI output that the source doesn't support: an invented finding, date, dose or event. In record summaries the riskiest kind is plausible, such as the result of a test that was ordered but never resulted. See failure modes.

Omission

A fact in the source that matters to the question but is missing from the summary. The most common error in the studies on the evidence page, and the 1 a read-through can't find.

Negation error

A negated finding stated as present, or the reverse: "denies numbness" summarized as "numbness". 30% of hallucinated sentences in the 2025 npj Digital Medicine study were negations.

Hedge

A word that marks clinical uncertainty: rule out, possible, probable, cannot exclude, suspicious for. A dropped hedge turns a differential into a diagnosis. Not to be confused with "ruled out", which means excluded.

Attribution error

A summary that changes who said something: the patient's report stated as a clinical finding, or a patient's account of another doctor's advice stated as that doctor's recommendation.

Copy-forward

EHR text carried from an earlier note into a later 1. A summary that trusts the later note can report a stale plan, exam or work status as current. In a 2022 study, 50.1% of note text at 1 health system was duplicated from earlier text.

Co-mingled records

Pages from another patient inside a record production, often with a similar name or the same facility and date. Any summarizer that reads them as your patient's adds their diagnoses to the history.

Laterality

Which side of the body a finding is on. Laterality errors (right for left) come from OCR, stray lines in a note or co-mingled pages.

Dates

Date of service

The date an encounter happened. The date to use for a visit in a summary, as opposed to fax, print, dictation or signature dates. See the 12 date types.

Addendum

Text added to a note after it was first written, with its own date. What an addendum says wasn't known at the original visit.

Citations and checks

Page citation

A pointer from a summary line to the source page it came from. It proves a page exists, not that the page supports the line. See citation checks.

Citation granularity

How precisely a citation points: to a whole document, to a page, or to a highlighted sentence. Finer granularity makes each check faster.

Upper bound

The highest error rate a random sample rules out at a stated confidence. 0 failures in 59 random lines gives a 95% upper bound of 5.0%. Try the sampling calculator.

Reverse spot check

Sampling the source instead of the summary: read random source pages and confirm their key facts appear in the summary. The sampling method that can find omissions.

Deduplication

Finding and setting aside repeated pages before a summary is built, so 3 copies of 1 visit don't read as 3 visits.

OCR

Optical character recognition: turning a scanned image into text. OCR errors on doses, sides and dates enter a summary before any language model sees the page.

Privacy terms

PHI

Protected health information under HIPAA. Real record files contain it; the synthetic test file on this site doesn't.

BAA

Business Associate Agreement: the HIPAA contract a vendor signs before it handles PHI for you. Sign it before uploading any real record to a summary tool.

Questions about these terms