What is OCR — and why AI misreads numbers more often than letters

You photograph a receipt and ask an AI to pull out the total. It does. Neatly, in the exact format you asked for.
And the number is wrong.
Not off by a cent — a different digit entirely. The nasty part: the mistake doesn't look like a mistake. This is OCR, and in 2026 it has picked up a rather sneaky new disease.
OCR turns pictures into text
OCR (Optical Character Recognition) is the conversion of an image of letters into real text you can copy and search.
To a computer, a photo of a document is just a grid of dots. The word "Total" on it is no different from a smudge: coloured pixels either way. OCR is the layer that says "these pixels are the letter T".
You use it every day without noticing:
- scanned PDFs that search suddenly works on;
- a sign translated through your phone camera;
- a bank pulling the amount out of a photo of a receipt;
- "photograph your card instead of typing it in".
Two generations of OCR — and they fail differently
This is the important part.
Classic OCR (Tesseract, PaddleOCR and friends) works like an engineer: it finds patches of text, cuts them into symbols, and for each one returns the closest matching character plus a confidence score. It's deterministic — the same file always gives the same result. If the pixels are blurred it outputs junk like 1t0r0 — but the junk is visible, and the confidence on those characters is low.
A multimodal model is an AI that "looks" at the image and generates text. It understands context: it knows "Total" and "Amount due" mean the same thing, and pulls the right field out of any layout with no setup. On formal metrics these models have been beating the classics for a while.
But it fails in a different way. When there aren't enough pixels, it doesn't stay quiet and doesn't emit junk — it writes something plausible. Not "couldn't read it", but a confident invention in place of the gap. That's exactly a hallucination, just in vision rather than in text.
Also: it has no honest "confidence in this pixel". A model can report a probability for its word, but that's not the same as "I can barely see this squiggle".
Why digits suffer more than letters
This isn't folklore, it's a measured effect. The DocHallu work (ACL 2026) deliberately damaged documents and watched where models started inventing. The finding: hallucinations occur across all models tested, and they happen more on numerical content than on text.
The logic is simple once you think about how a model works.
A blurred word gets reconstructed by meaning — the same way you read through typos without stumbling. Language is redundant: "Tot_l" resolves to exactly one thing.
A number has no redundancy. In the total 1 8 4 0, any digit in any position is equally plausible. Swapping 8 for 3 breaks no grammar, breaks no meaning, leaves the sentence perfectly fluent. There's nothing to check it against.
That's what makes the numeric failure the dangerous one: it leaves no trace. A wrong word looks off; a wrong total looks like a total.
How to use this without getting burned
You don't need to go back to classic OCR. You need to stop trusting digits without verification.
- Give it a decent image. Models invent most where visibility is poor: blur, shadow, low contrast. Reshoot straight and in good light and half the problem disappears.
- Ask for a strict output format, not prose. JSON fields are far easier to check than a paragraph.
- Give it permission to say "I can't see it". Literally: "if a character is unreadable, return
null, don't guess." By default a model tries to be helpful rather than honest. - Check numbers with arithmetic, not with your eyes. Line items should add up to the total, a date should land in a sane range, an account number should pass its checksum. That's thirty lines of code catching exactly the class of error eyes can't catch.
- Run it twice on anything that matters. The model is stochastic: the same file can give a slightly different answer. Two matching runs is reassurance. A mismatch is your "visibility is poor here" signal.
How is OCR different from computer vision?
OCR is one narrow task inside computer vision: extract the text. Computer vision is broader — "what is in this picture at all": faces, objects, defects on a part. Modern multimodal models do both, but the tasks are different.
Can OCR read handwriting?
Reasonably well now — much better than classic engines, which essentially gave up on cursive. But this is where the gap between "read it" and "made it up" is widest: handwriting is ambiguous, and the model is obliged to write something. Always double-check handwritten numbers.
Do I have to pay for OCR?
No, you can start free: Tesseract installs locally and runs offline, and most multimodal models have a free tier. Paying starts to make sense when you need volume and consistency, not recognition as such.
The AI Architect 2027 course: step by step from computer basics to your own app and first customers. 543 short lessons on your phone.





