How to send a PDF to ChatGPT

One thing is worth understanding before the steps, because it explains most of what follows. ChatGPT does not open a PDF the way a PDF reader does. It pulls the digital text out of the file and works from that, discarding the images on every plan except Enterprise. A PDF is a container, and what matters is whether there is real text inside it.

Sending a PDF on a computer

1. Open a chat

   A new conversation is cleaner than an old one. Everything already in a chat stays in context and competes with your document for attention.

2. Attach the file

   Click the paperclip or plus icon at the left of the message box and pick the PDF. Dragging the file straight onto the chat window does exactly the same thing and is faster.

3. Wait for the tile to settle

   A tile with the file name appears above the message box. Until it stops showing a loading state the file is not in yet, and sending early sends your question without the document.

4. Ask in the same message

   Type what you want done before you press enter, rather than sending the file on its own.

Sending a PDF on a phone

The apps handle uploads too, and the limits are identical, because they belong to your account rather than your device.

1. Tap the plus button

   It sits at the left of the message box in both the iOS and Android apps, and opens a menu with camera, photos and files.

2. Choose Files, not Photos

   Files reaches your device storage and any cloud drives you have connected. Photos only offers images, and a photograph of a page is not the same thing as the PDF.

3. Get the PDF into Files first if you need to

   Most PDFs arrive on a phone as email attachments. Open the attachment, use the share sheet to save it to Files or your drive, then attach it from there.

The camera is genuinely useful for a page or two, since ChatGPT reads images directly. It is the wrong tool for a long document: photographed pages go in as pictures, not text, and you will be fighting a much lower ceiling.

Then check it actually read the thing

This is the step almost everyone skips, and it is the difference between an answer you can trust and one that only sounds right.

Ask something that can only be answered from the end of the document. "What is the final section called?" or "quote the last sentence" will do. If it cannot tell you, it did not get that far, and whatever it just summarised was the opening pages.

If it says the PDF is empty or has no readable text

You have a scan. This is the most common confusion with PDFs and it is not a bug.

Retrieval from documents is text only, except Enterprise. ChatGPT pulls the digital text out of a file and discards any images. A scanned page is a photograph, so there is no digital text in it and the extraction comes back with nothing.

The test takes two seconds: open the PDF and try to select a sentence with your cursor. If you get a box drawn over the page instead of highlighted words, it is an image.

The fix is OCR, which turns the picture of the words back into words. Preview on Mac and Acrobat's Recognise Text both do it, and there are free web tools if the document is not confidential. More detail in why ChatGPT cannot read a scanned PDF.

If only the first part got read

This one has no error message, which is why it catches people. The limit that matters is 2 million tokens per text or document file, not the 512 MB size ceiling. A three-hundred-page report can be a couple of megabytes and still be well past what fits.

When a document is over that line it is truncated, not rejected. The upload looks perfectly normal and ChatGPT answers from as much as it could hold.

Three things work, in order of effort:

• Send only the section you need. Extract those pages into their own PDF. You usually get a better answer from twenty relevant pages than from three hundred mixed ones.
• Split it and send the parts in order. Tell ChatGPT first that a long document is coming in several parts and that it should reply "received" and wait rather than summarise. Without that instruction it answers part one as though it were the whole thing.
• Use a retrieval tool for anything book length. Past a point, even ordered chunks fall out of memory before the last one lands.

If the upload will not start at all

A refusal is the easiest failure to deal with, because it tells you something.

Retrying an upload that failed for size or length will fail identically every time. Zipping the PDF does not help either, since ChatGPT does not open archives.

The short version

Paperclip on desktop, plus button on mobile, question in the same message. That is the whole procedure and it works for most documents first time.

If the answer looks thin, work out which of the three it is. Nothing readable means a scan, so run OCR. A summary that stops early means truncation, so send the section you need or split it properly. A refusal means a size or plan limit, and the error will say which.