Can ChatGPT Read PDFs? What Works and What Does Not

Yes, and the reliability depends entirely on which of two mechanisms your file triggers. A typed PDF carries a text layer, and ChatGPT parses those characters directly, which works well. A scanned or photographed PDF holds no text layer, so the model reads the page as an image instead, and that path degrades quietly.

Which one you have is not a guess. Open the file and drag your cursor across a line of text. If the words highlight, there is a text layer underneath and extraction will be clean. If nothing selects, the page is a picture of text, and every warning below about the second mechanism is aimed at you. That one test decides more about the quality of your study material than any prompt you write afterwards.

How does ChatGPT read a PDF file?

Two mechanisms sit behind the single word read, and they are not variations of the same thing.

The first is text extraction. Most PDFs you meet at university are born digital: a slide deck exported to PDF, a Word document saved as PDF, a journal article, a chapter from an ebook. Files like these store the actual characters, and a PDF reader hands those characters straight over. Nothing is being interpreted or guessed at that stage, which is exactly why the path is dependable. What comes back is what the file already held.

The second is visual reading. A page that went through a library scanner, or was photographed on a phone, contains no characters at all. It contains an image of characters, which is a different kind of object entirely. To do anything with it, the model has to look at the picture and work out what the marks say, which is the same problem your eyes solve when reading handwriting that is not your own.

Everything frustrating about PDFs comes from the second path being a genuine interpretation while the first is a straight copy. The same chatbot, the same upload button, and two completely different degrees of confidence in the result.

Does ChatGPT read scanned PDFs and handwriting?

Often, yes, and the fair framing is that it can attempt what a plain text extractor cannot attempt at all. Printed text in a clean, straight, well-lit scan usually comes through in good shape. Four things reliably push it downhill.

• Handwriting, and cursive most of all. Joined letters carry individual quirks, and a page of lecture scrawl written at speed is the hardest case there is.
• Faint or skewed pages. A pale photocopy, a scan captured at an angle, or a phone photo with a shadow across the middle all cost accuracy.
• Dense tables running over several pages. Every cell has to be located as well as read, and the more crowded the grid, the more places there are to slip.
• Low-resolution page images. Detail that was never captured cannot be recovered later, whatever the model is asked to do with it.

The shape of the failure matters far more than the rate of it, and this is the part students underestimate. Errors from visual reading do not announce themselves. You do not get a gap, a warning, or a garbled string that has obviously gone wrong. You get a 3 where the page said 8, a table row that silently vanished, or a word misread into a different real word that still fits the sentence. The output arrives as ordinary competent prose either way, which removes the exact cue you would otherwise use to catch it.

So the rule is narrow and cheap to follow. Anything numeric, anything you would quote, and anything a mark depends on gets checked against the original page. The rest can be taken on trust.

How large a PDF will ChatGPT accept?

The published ceilings are 512MB per file, plus a separate cap of 2 million tokens for text and document files, both set out on the OpenAI Help Center file uploads FAQ. Those are the two numbers with a named owner behind them, so they are the two worth remembering.

Neither is usually what stops a student. A typed lecture PDF runs to a few megabytes at most, and a whole module of slides is often smaller than a single photograph. Files approaching 512MB are essentially always image-heavy, which drops them straight back into the scanned category and its reliability problems. The token ceiling is the more interesting constraint, because it governs how much text can enter the conversation from one document rather than how many bytes the upload weighs.

Two things you will read elsewhere deserve a caveat. Per-plan upload frequency figures circulate widely, along the lines of a small daily allowance on the free tier and a much larger rolling one on paid tiers. Those numbers come from third-party trackers rather than any OpenAI-published page, so treat them as directional at best. And no official page-count limit exists at all, so a post quoting one is describing whatever happened to work for its author on the day.

The other question people ask at the upload button is not how large the file may be but what becomes of it once it lands, which is answered separately in the walkthrough of what happens to notes you upload to ChatGPT.

What happens to tables and formatting?

Formatting is the first casualty on both paths, and it fails in a way that is easy to miss.

Where a table sits in a real text layer, the numbers themselves generally survive, because they are being copied rather than interpreted. The arrangement is what gets lost. Rows and columns flatten into running text, headers detach from the values beneath them, and a wide table or one that breaks across a page can come back with figures correctly transcribed and wrongly attached. Where the table is part of a scanned image, both the values and the layout are exposed, since every cell has to be found before it can be read.

Two habits cover nearly all of it. Ask for the table restated as a list of labelled rows rather than as a grid, because a dropped or misaligned row is obvious in a list and invisible inside a reflowed block of text. Then verify any figure a question actually turns on. That is the same discipline that applies to how far ChatGPT can be trusted on specifics in general, and it costs seconds rather than minutes.

Reading a PDF and studying from it are different jobs

A model that has read your chapter can answer questions about it. That is genuinely useful, and it is not the same as learning the chapter. The gap shows up at the point most people stop: a clean summary, read once, feels like progress and leaves almost nothing behind a week later.

What converts a read PDF into retained material is being asked about it without the document in front of you, repeatedly, until the recall belongs to you. The mechanics of turning a chapter into that kind of material are covered in the guide to making flashcards from a PDF. If you would rather interrogate a document conversationally than generate questions from it, the roundup of chat-with-PDF tools built for students compares the products aimed at exactly that.

Does GeniusPal read scanned PDFs?

No, and that limitation deserves stating as plainly as the strength sitting next to it. GeniusPal lands squarely on one side of the split above. Extraction runs through unpdf, a library built on pdf.js, and it pulls the text layer. That is the entire mechanism. No vision-based fallback exists anywhere in the pipeline, so the reliable half of what ChatGPT does is matched exactly, and the other half is never attempted.

The consequence is easy to state and worth stating rather than glossing. A typed or digitally exported PDF, which covers slide decks, Word documents saved as PDF, and most textbook chapters, extracts cleanly and turns into a usable set of questions. A scan or a photograph of handwritten notes yields little or no usable text, and retrying changes nothing, because there is nothing in the file for a text extractor to find. If that is what you have, type the notes up or run them through a dedicated OCR tool first, then upload the result. ChatGPT will make a partial attempt at the same file and may recover some of it, with all the caveats above attached.

The size ceiling is a smaller story than it appears. GeniusPal accepts up to 10MB against the 512MB ChatGPT allows, which sounds like a wide gap and rarely behaves as one: a term of typed lecture notes or an exported slide deck usually sits well under a few megabytes. Files that run past 10MB are almost always image-heavy, which is the scanned case that would extract badly regardless of any ceiling. Alongside PDF, uploads accept Word, PowerPoint, plain text, Markdown, and CSV.

What comes back deserves the same precision, since expectations here get set by other products. GeniusPal turns an upload into a question set: a quiz, flashcards, or an active-recall drill. It does not produce a mind map or a summary, and it is not a chat window you can question about the document. The free tier is two study-set generations for the lifetime of the account rather than a monthly allowance, which is enough to find out whether questions written from your own material are any good.

So the short answer holds with one condition attached to it. ChatGPT can read PDFs, and how well it reads yours was settled before you uploaded anything, by whether the file holds text or holds pictures of text. Check that first, fix it when you need to, and every tool downstream, this one included, gets a fair run at the material.