Glossary
OCRoptical character recognition
OCR, or optical character recognition, is technology that reads the letters and numbers inside an image of a page and converts them into text a computer can search, copy and edit.
In plain terms
A photograph of a receipt is, to software, just a picture. It has no idea there is a total on it. OCR is the step that looks at the shapes in that picture and works out that they spell words and figures. Once that has happened, the receipt behaves like anything else you have typed: you can search it, copy from it, or hand it to something else to process.
Why it matters
A great deal of the paperwork that matters most arrives as pictures. Scanned contracts, photographed invoices, screenshots, forms returned by post, decades of filed documents. Until the text inside them is readable, none of it can be searched, summarised or fed into any other system, and the work falls to whoever is willing to retype it. OCR is the unglamorous step that turns that pile into something the rest of your tools can act on, and it is usually the first thing to check when a document workflow refuses to work.
How it works
The classic approach works in stages. The image is cleaned up first: straightened, contrast adjusted, background speckle removed. The layout is then analysed to separate columns, paragraphs, tables and images from one another, because reading a two-column page in the wrong order produces nonsense. Only then are the individual characters recognised and assembled back into words, with a dictionary check catching many of the misreads.
Newer systems fold most of those stages into a single trained model that has seen very large numbers of document images. This handles the hard cases better: handwriting, receipts creased in a pocket, photographs taken at an angle in poor light, pages mixing several languages. Multimodal assistants take this further again, describing and answering questions about a page in one step rather than producing a text file you then have to read.
Layout is where the difficulty concentrates. Getting the characters right is largely a solved problem on clean printed pages. Preserving the structure that gives them meaning is not: which figure belongs to which column of a table, which line is a heading, which of the four amounts on an invoice is the one you actually want. Anything that reads documents for a downstream process is being judged on structure at least as much as on characters.
The output is therefore a confidence estimate rather than a certainty. Good systems report how sure they are, and the sensible design puts a person on the low-confidence cases rather than on all of them.
What a reader gets back, and where it goes wrong
Seen in the wild
Photograph a page of a printed document with your phone, give it to a general assistant that accepts images, and ask for the text back as something you can edit.
ChatGPTUpload a scanned agreement to an assistant built for long documents and ask what it commits you to, then check the answer against the page it came from.
ClaudeGive a screenshot of a table to a multimodal assistant and ask for it back as rows and columns, which is the layout half of the problem rather than the character half.
Google Gemini
Common misconceptions
People assume
A PDF is already text, so OCR is not needed.
In fact
Some PDFs are text and some are pictures of pages wrapped in a PDF. If you cannot select a sentence with your cursor, there is no text in there and something has to read it first. This is a frequent reason a document tool gives an empty or useless answer about a file that looks perfectly readable to you.
People assume
OCR is accurate enough to trust unchecked.
In fact
On clean printed pages it is very good. On handwriting, faint thermal receipts, stamps, poor photographs and dense tables it still makes mistakes, and the mistakes that matter are usually digits rather than letters. Anything financial or legal warrants a check of the specific figures you are relying on, not the whole document.
Telling them apart
OCR vs Computer vision
OCR
Concerned with one question: what does this text say. The output is characters, words and, at its best, the structure holding them.
Concerned with what is in an image generally: objects, faces, damage, movement, counts. Text is one of many things it might be asked about.
OCR is a specific job within computer vision. If the answer you want is words, it is OCR. If the answer is what is happening in this picture, it is the wider field.
Questions
- What does OCR stand for?
- Optical character recognition. Optical because it works from an image, character because it identifies individual letters, digits and punctuation, and recognition because it matches shapes against known forms. The name dates from long before AI, and the modern versions work quite differently underneath while answering the same question.
- Can OCR read handwriting?
- Modern systems often can, with far more success than older ones, but accuracy varies with the handwriting and drops sharply on cursive, cramped or inconsistent script. Treat handwriting as a case that needs checking rather than one that needs trusting, and keep a person on anything where a misread digit would be expensive.
- Why did my document tool return nothing useful from a scan?
- Most often because the file is an image with no text layer, and the tool was never given a reading step. Try selecting text in the file: if the cursor will not highlight a sentence, nothing in there is text. Running the scan through a tool that reads images first usually fixes it.
- Is OCR the same as document AI?
- No. OCR produces the text. Document AI is what comes after: deciding which number is the invoice total, which date is the due date, and routing the result into a system. Many products bundle the two, so the distinction only surfaces when one half works and the other does not.
Key takeaways
- OCR turns a picture of a page into text that can be searched, copied and processed.
- A PDF you cannot select text in is an image, and needs reading before any tool can use it.
- Recognising characters is largely solved; preserving layout and tables is the hard part.
- Accuracy falls on handwriting, poor photographs and creased receipts, and misread digits are the costly failure.
- It is the first step in a document workflow, not the whole of one.
Tools that use this
- ChatGPT
Accepts photographs and documents, so a picture of a page can be turned back into text.
- Claude
Handles long uploaded documents, which is where reading a scan usually matters.
- Google Gemini
Understands images and files directly, which covers the layout half of the problem.
Last checked July 2026