Guide

What OCR gets wrong, and how to catch it

Text recognition does not read — it guesses, character by character, and the guesses are usually right. That is exactly why the wrong ones are dangerous: they arrive in fluent, confident text with nothing marking them.

Updated November 2026 · 7 min read

A guess that looks like a copy

When you photograph a page and run text recognition on it, what comes back looks like the page. Paragraphs break in the right places, sentences read correctly, punctuation is present. The presentation is so faithful that it invites trust.

Internally, though, nothing was read. The system looked at shapes and asked, at every position, which character most likely produces that shape in that context. Most of the time the answer is obvious and correct. Occasionally it is a coin flip — and a coin flip that produces a plausible word leaves no trace.

That is the whole problem in one sentence: recognition errors do not look like errors. They look like text.

Where the mistakes cluster

Failures are not evenly distributed. They concentrate in a handful of places, which makes them predictable enough to check.

Numerals

Digits carry almost no surrounding context to help disambiguate them. In a sentence, "cat" and "cot" are separated by the rest of the sentence; in a table cell, "8" and "6" are separated by nothing. Currency amounts, page numbers, measurements, and dates are where OCR errors do the most damage and are least likely to be noticed.

Lookalike characters

These cause substitutions that survive spellcheck because the result is still a real word or a plausible string:

Looks likeOften becomesTypical victim
0 / Oeach othercodes, model numbers
1 / l / Ieach otheridentifiers, units
rn / meach othernames, small text
5 / Seach otheramounts in small print
2 / Zeach otherreferences

Small, tight, or stylised type

Footnotes, captions, legal text, and anything set in a condensed or decorative face are recognised less reliably than body copy — and they are usually the parts that matter most when you quote something.

Rotated and skewed text

A few degrees of tilt is enough to degrade recognition noticeably. A page photographed at an angle, or a label photographed sideways, is a common source of garbled output.

Mixed languages and diacritics

Accents, umlauts, and non-Latin scripts are recognised with lower confidence, and switching language mid-page — a quotation inside an English sentence, a technical term in another script — is where models most often drop or replace characters.

Layout is a separate problem

Recognition returns characters in an order, and that order is a guess too. Single-column body text comes out almost perfectly. Multi-column pages sometimes interleave the columns, producing sentences that are grammatically fine but read across two unrelated paragraphs.

Tables are the worst case: cells are frequently merged into a single run of text, so the relationship between a row label and its value is lost. Footnotes get attached to the wrong paragraph, or moved to the end of a page where their reference number no longer means anything.

If the result reads oddly but every word is spelled correctly, the problem is layout, not recognition — and re-recognising will not fix it. Retype the structure by hand.

Getting a better input

Most recognition failures are decided before the image reaches the model. A few habits improve the result more than any setting:

  • Crop tight. Include the text you want and as little else as possible. Margins, table borders, and neighbouring columns are noise.
  • Straighten it. Hold the camera square to the page. On a screenshot, make sure the text is not scaled to the point where strokes blur together.
  • Increase contrast. Light grey text on white, or text over a photograph, is the hardest case. A crop without the background helps more than any enhancement filter.
  • Avoid glare and shadow. A bright reflection across a page removes the strokes it covers; nothing downstream recovers them.
  • Prefer the original. If you are photographing a screen, take a screenshot instead. Screen text is pixel-perfect and recognition on it is close to reliable.

The verification pass

Checking every character is unrealistic. Checking the parts that matter takes under a minute, and it follows a fixed order:

  1. Every number. Compare against the original one at a time. Prices, quantities, dates, versions, percentages, page references.
  2. Every name. People, companies, products, places. Proper nouns are where a substitution is least likely to look wrong and most likely to matter.
  3. Every quotation mark. Confirm the quoted passage begins and ends where the original does — recognition sometimes swallows an opening mark or promotes a stray apostrophe.
  4. The first and last sentence. Edges of an image are where text is clipped or half-recognised.

If the text will be published, cited, or used to make a decision, add the original image next to it. The image is the evidence; the text is the working copy.

When to retype instead

Recognition is not always worth it. If the passage is under about three sentences, typing it is usually faster than recognising, reading, and correcting — and the result is exact. The same applies to anything containing several numbers, any table you need to preserve as a table, and any text in a script or with diacritics you can type yourself.

Recognition earns its place when the volume is high and the content is ordinary prose: a scanned letter, a photographed page of a book, a slide of bullet points, a sign with no numbers on it.

How Papscroll handles it

Papscroll's area capture sends the region you select to a vision model and returns plain text for you to review before saving. It is designed for the moment when retyping is the only alternative, not as a replacement for copying text that is already text.

The returned text is editable, and the note keeps the rest of its context — title, source, and your own remarks — so a recognition result never stands alone as an unsourced block of words.

A short rule that prevents most of the damage: treat recognised text as a draft. It is a very good draft, but it is a draft.