Why ChatGPT can't read your PDF (and how to fix it)
You attach a PDF, ask a question, and the answer is wrong, vague, or "I can't see any text in this file". It's rarely the AI being difficult. Usually something about the PDF makes it hard to read. Here are the common causes and how to fix each one.
1. The PDF is a scan
Symptom: the assistant says the document is empty, only describes the page, or makes up content.
A scanned PDF is a set of photos of paper. There's no text layer inside, only pixels. Some assistants can read page images, but results vary, especially for small print, stamps, handwriting and non-English scripts.
Fix: run OCR first and give the assistant the text. Drop the PDF on the Docuflat converter, choose the document's language, and paste the Markdown into the chat. You can check the text yourself before you rely on the answer. See How to convert a scanned PDF to text.
How to tell: try selecting a sentence in your PDF viewer. If you can't highlight individual words, it's a scan.
2. The file is too long
Symptom: answers about the start of the document are fine, but questions about later pages are wrong or ignored.
Every assistant has a limit on how much it can take in at once (its "context window"), and long PDFs can exceed it or crowd out the rest of the conversation. Some tools also only read part of a large file.
Fix: convert to Markdown, which is much smaller than the PDF, and send only the sections you need. Our guide on giving a very long document to an AI explains how to split it sensibly.
3. The text comes out garbled
Symptom: the assistant quotes nonsense characters, or Hindi, Marathi or other text appears as random English letters.
Some PDFs, especially older Indian-language documents made with legacy fonts, store text in a private encoding. It looks right on screen, but the underlying characters are wrong, so anything that reads the text, including AI, gets gibberish.
Fix: treat the page as an image. Take a screenshot or export the pages as images, then convert them with OCR in the right language. The OCR reads what's visible, not the broken encoding. More in How to convert a Hindi PDF to text.
4. The PDF is password-protected or restricted
Symptom: the upload fails, or the assistant says it can't open the file.
Fix: open it with the password and save or print an unprotected copy (only for documents you're entitled to do this with). Docuflat also asks you to remove the password before converting.
5. Tables and columns are mixed up
Symptom: numbers end up in the wrong rows, or the answer mixes text from two columns.
PDFs store positions, not structure, so multi-column pages and tables can be read in the wrong order.
Fix: convert the PDF, check the order on the pages that matter, and rebuild key tables as Markdown tables. If you have the original spreadsheet or Word file, use that instead. See How to extract a table from a PDF.
6. You've hit a usage limit
Symptom: the assistant stops partway, or you reach your message limit far sooner than usual.
PDFs can use many more tokens than their text alone, especially when pages are processed as images too.
Fix: paste Markdown instead of attaching the PDF. It's the same information for a fraction of the tokens. See How to use fewer tokens with documents.
A routine that avoids most problems
- Convert the PDF to Markdown, with OCR for scanned pages.
- Skim the result: is the text readable, in the right order, complete?
- Remove pages you don't need, such as covers, indexes and boilerplate.
- Paste it into the chat and tell the assistant what the document is and what you want from it.
You'll get better answers, use less of your allowance, and know exactly what the AI was given.