How to upload a scanned document to ChatGPT

The uploading is not the problem. A scanned PDF uploads perfectly happily, and then ChatGPT says it cannot find any text, or answers vaguely from the file name. Retrieval from documents is text only, except Enterprise, so a page that holds only an image leaves nothing behind.

OCR, optical character recognition, turns the picture of the words back into words. Everything below is that one step, done properly and then checked.

Confirm it is really a scan

Three tells, fastest first.

Try to select a sentence. Click and drag across some words. If they highlight one at a time, following your cursor, there is real text and your problem is something else. If a rectangle draws over the page instead, it is an image.

Search the file. Press Cmd+F or Ctrl+F and look for a word you can see on the screen. No matches means no text layer.

Glance at the size. Forty pages of text is usually a few hundred kilobytes. Forty scanned pages run to several megabytes, because each page is a photograph. Not proof on its own, but a good hint.

Documents are often mixed. A report written digitally with a scanned appendix passes the selection test on page one and fails on the page you care about. Test the section you actually want to ask about.

OCR on a Mac

Preview, for a few pages. macOS recognises text inside images itself. Open the scan in Preview and try to select a line. If the cursor picks out words, you can copy them straight into a chat, with nothing installed.

Preview export, for a text layer. File, then Export as PDF. On recent macOS versions the recognised text is written into the exported file. Run the selection test on the export rather than assuming, since this varies by version.

Shortcuts, for a folder of images. The Shortcuts app has an Extract Text from Image action. Point it at a folder of page images and it writes out the text in one pass.

OCR on Windows

PowerToys Text Extractor. PowerToys is free from Microsoft. Text Extractor copies text out of any region of the screen, including a PDF page you happen to be looking at. No conversion, no upload, good for a handful of pages.

OneNote. Insert the image into a page, right click it, and choose Copy Text from Picture. It has been there for years and it is reliable on clean scans.

Not Word. Word will open a PDF and convert it, but the pages of a scan usually arrive as pictures in the converted document. It is not an OCR step.

For a whole document, or a private one

OCRmyPDF. Free and open source, built on Tesseract. It adds an invisible text layer under the original page images, so the document looks identical and is now readable. It runs on your own machine and handles hundreds of pages. For anything long or confidential this is the right answer. It is a command line tool, which is the only real cost.

Adobe Acrobat. Scan and OCR, then Recognise Text. Paid, and the most forgiving on poor scans.

Google Drive. Upload the file, right click it, open it with Google Docs. OCR runs and you get a document you can copy from. Layout is flattened and long files do not always come through whole, so check the last page is there.

Free OCR websites. Most of them work fine.

After OCR it is an ordinary document

Which means the ordinary limits apply: 512 MB per file, and 2 million tokens for text and document files.

The size cap rarely bites, but a scan is one of the few cases where it can, because an OCR'd PDF still carries a full page image for every page. If you are close to it, export the text on its own as a .txt or .docx and the file collapses to a fraction of the size. You lose the page images, which retrieval was discarding anyway.

The length cap is the more likely wall. A scanned book that comes through OCR cleanly can still be too long to be read in one piece, so you go straight from one problem to the next.

Check the OCR worked before you upload

1. Run the selection test again

   On the new file. Words highlight, you have text. Still a rectangle, the OCR did not take and the upload will fail exactly as before.

2. Search for a word from the last page

   The step people skip. It confirms the text layer covers the whole document rather than stopping partway through.

3. Read a paragraph of the output

   OCR on a faded or skewed scan produces plausible nonsense, not obvious breakage.

4. Spot check the numbers

   Digits are where OCR is weakest and where a wrong answer costs most. Zero and O, one and l, and decimal points that quietly vanish. Check any figure you plan to ask about.

5. Once it is uploaded, ask what the final section is called

   If ChatGPT cannot say, the text layer is incomplete and you are getting answers about part of the document.

Bad OCR does not fail loudly. A document full of mangled words still uploads, still gets read, and still produces a confident answer. You cannot tell it is wrong.

When to skip OCR entirely

A few pages. Screenshot them and upload the images. Images are supported, capped at 20 MB each, and quicker than any OCR setup for three pages. Impractical for three hundred.

Handwriting. Recognition built for print is unreliable on handwriting. Type out the part you need instead.

A bad scan. Skewed, faded, or photographed at an angle. Rescanning flat, in even light, at a decent resolution beats any amount of cleanup afterwards.

ChatGPT Enterprise. Enterprise supports visual retrieval for PDFs, so a scan may work as it is. Every other plan needs the text layer.

The short version

Selection test first. If a rectangle draws over the page, OCR it with whatever you already have: Preview on a Mac, PowerToys or OneNote on Windows, OCRmyPDF for anything long or private. Then test the result before you upload. A silent OCR failure costs far more than the minute it takes to check.