A scanned document that reads perfectly can still be wrong in the one place it counts.
Modern OCR is good — startlingly good on clean print. Point a decent scanner app at a lease, a lab report, or a utility bill, and it will hand you searchable text with nearly every word correct. But "nearly every word" is a strange kind of comfort, because OCR errors are not scattered evenly across a page. They pool, quietly, in the places language can't protect: account numbers, dosages, dates, reference codes, names. And the proofreader of last resort — your own reading brain — is beautifully engineered to miss exactly those.
This article is about closing that gap. Not by proofreading harder, which doesn't work, but by understanding where machine errors and human blind spots overlap, and checking only the narrow territory where they do.
The errors worth worrying about aren't in the words
Here is the asymmetry at the heart of the problem. If OCR turns the into tbe, almost nothing happens. You notice it instantly, or a search still finds the sentence, or the meaning survives anyway. Prose is redundant by nature — English carries so much internal structure that a mangled word rarely stays hidden and rarely does damage.
Numbers have no such safety net. There is no dictionary of valid phone numbers, no grammar of invoice totals. If a scanner reads a 4 as a 1, the result isn't gibberish that flags itself — it's a different, perfectly plausible number. An amount of 1,400 and an amount of 1,900 are equally fluent. A meter reading, a policy number, a medication strength: each is just a string of digits in which every digit is load-bearing and no digit can vouch for its neighbors.
Names sit in the same brittle category. Meiers and Meirs both look like surnames. Nothing about the text itself tells you which one was on the paper.
So the cost of an error and the detectability of an error run in opposite directions. Word errors are common in bad scans but cheap and conspicuous. Digit and name errors are rarer but expensive and nearly invisible. Any sensible checking strategy follows the cost, not the frequency.
You don't read letters — you read guesses
Why can't you just reread the document carefully? Because skilled reading is not inspection. It's prediction.
Psychologists have known for decades that fluent readers don't process text letter by letter. One classic demonstration is the word superiority effect: people identify a single letter more accurately when it appears inside a real word than when it appears alone or inside a nonsense string. That result is strange if reading were a bottom-up pipeline from letters to words — the word is somehow helping you see its own parts. It makes sense only if recognition runs top-down: your brain uses the whole word, and the sentence around it, to decide what the letters must have been.
This is why everyone has sailed past a doubled word — the the — when it straddles a line break, and why typos in familiar text are so hard to spot. You aren't seeing the page; you're seeing your model of the page, lightly corrected by glances. When you proofread the OCR of a document you already know — your own lease, your own invoice — prediction gets even stronger. You read what should be there.
Digits are the worst case for this machinery, from both directions at once. Your brain can't predict them, so it doesn't linger on them the way it puzzles over a broken word. But it also gets no error signal from them, because a wrong digit produces no ungrammatical, attention-grabbing wreckage. The eye slides over 7,215 exactly as smoothly as it slides over 7,275. The stakes are highest precisely where attention is weakest.
Where OCR actually stumbles
Knowing how recognition fails tells you what to look for. OCR errors are overwhelmingly substitutions between lookalike shapes: 0 and O, 1 and l and I, 5 and S, 8 and B, the pair rn fusing into m, cl collapsing into d. Degraded sources make everything worse — thermal receipts gone grey, old faxes, carbon copies, small print photographed at an angle in dim light. On a clean, well-lit, squarely captured page, substitutions are rare. On a faded receipt, they're routine.
Modern OCR engines add a layer that's easy to misunderstand: they lean on dictionaries and language models to resolve ambiguous shapes. A smudged glyph that could be o or a gets snapped to whichever spelling makes a real word. This is why today's output reads so cleanly — and it's a genuine improvement for prose. But notice what it does to your brittle fields. Language knowledge rescues words and does nothing for digit strings, because no model knows what your account number is supposed to be. Worse, the overall fluency of the output lowers your guard. A page with no visible garbage feels verified. It isn't. Fluency is not fidelity.
A verification habit that fits in five minutes
The practical conclusion is not "proofread everything." It's triage. Most of a scanned page never needs to be exact — you'll read it, not retype it. Check only the parts you'd someday need to reproduce character-for-character.
First, mark the brittle fields. Scan your eye down the document and pick out anything you might one day copy exactly: amounts, reference and policy numbers, dates, dosages, addresses, email addresses, proper names. On most documents that's a handful of items, not a page of them.
Second, verify against the image, never from memory. Put the original picture and the recognized text side by side and read from the pixels to the text, one field at a time. For long numbers, defeat your brain's autopilot deliberately: read the digits aloud in pairs, or read the string backwards. Both tricks strip away the fluent, predictive mode of reading and force the slow, letter-by-letter mode that proofreading actually requires. It feels pedantic. That's the point — the pedantic mode is the one that sees.
Third, let checksums work for you. Some numbers can validate themselves. Credit and debit card numbers carry a built-in check (the Luhn algorithm) designed to catch any single-digit mistake; IBANs include check digits that fail if a character was misread. If a scanned card or account number passes a free online validator, a lone OCR substitution is very unlikely to be hiding in it. Dates offer a softer version of the same trick: if the letter says "Tuesday, March 12," a calendar will tell you whether March 12 was a Tuesday.
Fourth, use search as a tripwire for names. Search the document for the name spelled the way you believe it should be. Zero results on a page where the name obviously appears means the OCR spelled it differently — go look at which version is right.
Keep the picture — it's the ground truth
There's one structural decision that matters more than any checking technique: never keep recognized text instead of the image. The best archival form for a scan is a searchable PDF, where the photograph of the page sits on top and an invisible text layer sits behind it. You search and copy the text; you read and trust the image. When a number becomes important eight months from now, you don't have to wonder whether the OCR was right — you zoom in on the actual ink.
Text-only exports have their uses, but for anything contractual, medical, or financial, the image is the document and the text is an index to it. Treat OCR output the way a careful editor treats a transcript: enormously convenient, and always one step removed from the source.
None of this takes long. Marking the brittle fields, checking them against the pixels, running a validator over the one number that self-checks — five minutes, usually less. It's the difference between a pile of text that is probably right and an archive you can act on without flinching.
This philosophy is built into LumenScan. Its OCR runs entirely on your device — no page, and no misread digit, ever leaves your phone — and every scan is saved as a searchable PDF with the original image kept intact behind the recognized text, so the ground truth you verify against is always one tap away. When you want text you can trust and a picture that can prove it, it's ready at https://lumenscan.lumenlabs.works.