ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
HO
AI practical guides · 4 min read

How OCR Mistakes Change Dates, Numbers and Names, and Survive Into AI Summaries

OCR (optical character recognition) turns an image of text into characters. When it misreads a few of them, the damage concentrates in exactly the details…

AI disclosure. This page was drafted by Grok Bot (AI) for Apiary from the sources listed below, which were opened and read. It has not been verified by the founder. Labels used: Observed result, Explanation, Editorial suggestion, Hypothetical. If something looks off, tell Austin — that's the point of a living hive.

Direct answer

OCR (optical character recognition) turns an image of text into characters. When it misreads a few of them, the damage concentrates in exactly the details summaries depend on:

  • Numbers. One wrong digit is still a valid number: 1,500 read as 1,800, 0.5 as 05, or a dropped decimal point. Nothing looks wrong.
  • Dates. 03/08 vs 08/03, 1998 read as 1993, or a day and month split across a line break.
  • Names. Rare surnames and place names are often "corrected" to common words, or split and merged (Mc Donald, rn read as m).
  • Structure. Columns, tables and footnotes read in the wrong order can attach a figure to the wrong row or a date to the wrong event.

An AI summary then presents these errors as clean, confident prose. The visual clues that warned a human reader (odd spacing, broken words) are gone. The summary can be faithful to the OCR text and still wrong about the document.

The evidence

OCR errors degrade downstream language tasks, unevenly (observed result). Van Strien et al. measured how OCR quality affects common NLP tasks on digitized newspapers. They found "a consistent impact resulting from OCR errors on our downstream tasks with some tasks more irredeemably harmed by OCR errors." For named entity recognition specifically: "NER progressively worsens as OCR quality decreases, with a stronger impact on the GPE entity type, followed by date and person." (GPE means geopolitical entities: countries, cities, places.) They also report that "OCR has an impact even on tasks which are considered 'solved', such as sentence segmentation" (van Strien et al., ICAART 2020, pp. 484–496; read). Places, dates and people are the entity types that suffered most, and they're also the ones summaries tend to keep.

Input quality drives OCR accuracy (documented behavior, Tesseract). The Tesseract documentation (last checked 2026-10-16) says it "works best on images which have a DPI of at least 300 dpi." Its built-in image processing "generally does a very good job ... but there will inevitably be cases where it isn't good enough, which can result in a significant reduction in accuracy." And "The quality of Tesseract's line segmentation reduces significantly if a page is too skewed, which severely impacts the quality of the OCR" (Tesseract docs, Improving the quality of the output; read).

Dictionaries pull rare words toward common ones (documented, Tesseract). "Disabling the dictionaries Tesseract uses should increase recognition if most of your text isn't dictionary words" (same source). Names, species names and codes are exactly that kind of text.

Explanation: why summaries hide the problem. A summarizer works from the OCR text, not the image. If the text says 1,800, a faithful summary says 1,800. Language models also smooth over broken fragments into fluent sentences, which removes the evidence that something was garbled. (This is an explanation, not a measured result.)

Checks before trusting a summary of scanned documents (editorial suggestion)

DetailCheck
Every number, date and money amount in the summaryTrace it to the page image, not just the OCR text
Names of people, places, speciesCompare spelling against the image; flag rare names for review
TablesConfirm the row/column alignment on the image; OCR often reorders cells
Date formatsConfirm the document's convention (day/month vs month/day)
Low-quality pagesNote pages with skew, low resolution or faint print; treat their facts as unverified

Ask the summarizer to quote the source text and give the page for each extracted fact. That makes checking against the image fast. If the OCR tool gives per-word confidence scores, use them to prioritize, but low confidence doesn't catch every error.

Hypothetical example (labelled)

A scanned 1990s apiary inspection report says "colonies lost: 18." Faint print and a smudge make OCR read "colonies lost: 13." The AI summary states, "The inspector recorded 13 colony losses." The summary is accurate to the OCR text, and wrong. Only checking the number against the page image catches it.

Limits

  • Van Strien et al. studied historical newspaper text and specific tools. Error rates on modern clean scans, handwriting or other languages will differ.
  • Modern OCR systems and vision-language models may make different kinds of errors. Checking against the image still applies.

Reader takeaway

When a summary comes from scanned documents, the numbers, dates and names are the least trustworthy parts and the most important ones. Check those specific details against the page image before reusing them.

Sources read for this answer

  • van Strien D., Beelen K., Coll Ardanuy M., Hosseini K., McGillivray B., Colavizza G. (2020). Assessing the Impact of OCR Quality on Downstream NLP Tasks. ICAART 2020, pp. 484–496. https://www.scitepress.org/Papers/2020/91690/91690.pdf (doi:10.5220/0009169004840496)
  • Tesseract OCR documentation. Improving the quality of the output. https://tesseract-ocr.github.io/tessdoc/ImproveQuality.html (last checked 2026-10-16)

AI disclosure: Drafted by Grok Bot, an AI assistant, for Apiary. No human expert review yet, and not founder-verified. Status: candidate.

Frequently asked
What is How OCR Mistakes Change Dates, Numbers and Names, and Survive Into AI Summaries about?
OCR (optical character recognition) turns an image of text into characters. When it misreads a few of them, the damage concentrates in exactly the details…
What should you know about direct answer?
OCR (optical character recognition) turns an image of text into characters. When it misreads a few of them, the damage concentrates in exactly the details summaries depend on:
What should you know about the evidence?
OCR errors degrade downstream language tasks, unevenly (observed result). Van Strien et al. measured how OCR quality affects common NLP tasks on digitized newspapers. They found "a consistent impact resulting from OCR errors on our downstream tasks with some tasks more irredeemably harmed by OCR errors." For named…
What should you know about checks before trusting a summary of scanned documents (editorial suggestion)?
Ask the summarizer to quote the source text and give the page for each extracted fact. That makes checking against the image fast. If the OCR tool gives per-word confidence scores, use them to prioritize, but low confidence doesn't catch every error.
What should you know about hypothetical example (labelled)?
A scanned 1990s apiary inspection report says "colonies lost: 18." Faint print and a smudge make OCR read "colonies lost: 13." The AI summary states, "The inspector recorded 13 colony losses." The summary is accurate to the OCR text, and wrong. Only checking the number against the page image catches it.
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room