Ambiguity in the glyphs themselves
Some characters are near-identical in many fonts. Zero and capital O, one and lowercase L and capital I, five and S, eight and B, and the pair r-n read as m are the classic confusions.
Context normally resolves these. In a word, a language model knows which is likely. In a serial number or licence plate there is no context, so these are precisely the places errors survive.
Damage introduced before recognition
Compression artefacts, motion blur, low resolution and shadows all remove information. No model recovers detail that is not present in the pixels.
Over-aggressive thresholding breaks thin strokes, turning e into c or a into o. Under-aggressive thresholding merges adjacent characters.
Layout errors that look like text errors
When columns are merged, every character can be correct while the document is meaningless. Read paragraph transitions, not just words.
Footnotes, marginal notes and captions are frequently spliced into the body text. Scan for sudden topic changes mid-paragraph.
A two-minute checking routine
First, search the output for digits and check each number against the original. Second, check proper nouns. Third, read the first and last line of each paragraph for continuity.
This catches the great majority of consequential errors without proofreading every word, and it scales to long documents.