Can ChatGPT Translate a Scanned PDF? What Works and What Doesn't

Yes, with a catch that trips up a lot of people. When you upload a PDF, ChatGPT works from the file's text layer: the digital text that a PDF exported from Word or a website carries. A scanned PDF has no text layer. Every page is a picture. OpenAI's help center says only ChatGPT Enterprise reads the images inside a PDF; on other plans ChatGPT extracts the file's digital text and skips its images. So on Free, Plus or Pro, a scan needs a workaround.

This guide covers both halves: how to get a scanned PDF translated in ChatGPT anyway, and what you get back. For a one-page letter the answer is a good translation. For a forty-page manual full of tables, it's a pile of text to rebuild or a rebuilt PDF to check line by line.

What ChatGPT does with a scanned PDF

On a personal plan, ChatGPT's first step with an uploaded PDF is to pull out its text, and a scan from a phone app or an office scanner usually has none to pull. What happens next isn't documented, and it varies. Sometimes ChatGPT writes and runs its own code to work through the pages. In our side-by-side test we asked it to translate scanned PDFs and keep the layout, and after a minute or two of work it returned translated PDFs. Because that route isn't a documented feature, treat it as a bonus, not something to rely on.

Images are a different matter. ChatGPT reads uploaded PNG and JPEG files with its vision model, and that is the documented way to have it look at a page. If your scan isn't being read, sending the pages as images is the fix.

How to translate a PDF with ChatGPT, step by step

1. 1Check whether the PDF has text. Open it and try to select a sentence. If individual words highlight, the file has a text layer and you can upload the PDF itself. If the whole page selects as one block, or nothing selects at all, it's a scan.
2. 2For a scan, upload page images instead. Export the pages as PNG or JPEG (most PDF viewers and scanner apps can) and attach those. Each image can be up to 20 MB. On the Free plan you get 3 file uploads a day, which a page-by-page document uses up at once; paid plans allow up to 80 files every 3 hours.
3. 3Ask for every line, not a summary. Spell it out: "Translate every line on these pages into English. Don't summarize or skip anything. Keep the page numbers, and keep table rows as table rows. Write [illegible] where you can't read something."
4. 4Work through it a few pages at a time. OpenAI's own advice for very large files is to split them or point the prompt at specific sections. Five to ten pages per message is far easier to check than fifty in one go.
5. 5Ask for a file if you need one. By default the translation comes back as chat text. Ask for a Word file or a PDF and ChatGPT can create one to download, but nothing promises the original layout survives.
6. 6Check it against the original. Compare names, dates, numbers and amounts with the source page. OpenAI itself warns that ChatGPT may not reliably extract exact values from scanned files and image-based tables.

When ChatGPT is enough

Plenty of real situations don't need a rebuilt document, and for those, chat AI is hard to beat:

• You received a one-page letter in a language you don't read and just need the gist before deciding what to do with it.
• You want to ask questions about a document, like "what deadline does this notice mention?", rather than read a full translation.
• You need a quick sanity check on a short scan before paying someone to translate it properly.
• The document is a few paragraphs of plain prose with no tables, forms or layout that matters.

For these, ChatGPT is quick and good at the job, and pretending otherwise would be dishonest. If all you have is a sentence or two, OpenAI's separate ChatGPT Translate page handles typed text in 40-plus languages.

Where it breaks down

Long documents hit the daily and per-chat limits

File size is rarely the problem: a document can be up to 512 MB and 2 million tokens. The limits that bite are the Free plan's 3 uploads a day and how much one conversation can keep in view. A thirty-page scanned contract usually means several rounds of uploading, prompting and pasting the answers together, while checking that nothing fell between the rounds.

Things go missing without a warning

In our test, a word vanished from a list on a simple page (miso dropped out of four fermented foods) and a section heading disappeared from a physics paper. Nothing in the output said so. On a long document the same kind of gap is much harder to spot, so anything where completeness matters needs a page-by-page check against the source.

Tables come back as text in chat

In a chat reply, a table comes back as a Markdown table at best and as running text at worst, with merged cells and column alignment lost. A customs declaration or a financial statement means something because of where each number sits, so that matters. When ChatGPT builds a PDF instead, it can do much better: in our test it rebuilt a financial table neatly. Check stamps, footnotes and margin notes either way, since they sit outside the main text.

The rebuilt page paints over the original

Ask ChatGPT to keep the layout and it may write code that lays a solid color over the original text and draws the translation on top. On plain backgrounds that looks fine. On a photo or a gradient the patch shows: in our test a grey box covered most of an aerial photo, and asking it to redo the page "as precisely as it can" brought back much the same result after seven minutes.

What happens to the file you upload

On Free, Plus and Pro, OpenAI may use your content to train its models unless you turn off "Improve the model for everyone" under Settings, then Data controls. Business, Enterprise and Edu accounts are excluded by default. Worth knowing before you upload a passport, a contract or a bank statement.

The document-out alternative

The gap ChatGPT leaves is the part a scanned-document translator is built for. Reglyph works on the page image itself: it reads the text, erases the original from the image and sets the translation back in the same place, so tables, stamps, figures and columns stay where they were, and what you download is a finished PDF. Every page in the upload is processed, up to 300 pages and 100 MB per file, and you can download the original and the translation side by side to check one against the other. It's machine translation too, with the same limits: clear printed scans work best, handwriting is left as written rather than translated, and official submissions still need a certified translator. Your first 5 pages are free.

A practical decision rule

Ask one question: does the output need to be a document?

• No. You need the meaning, a summary, or an answer to a question: use ChatGPT. It's fast, conversational and good at exactly this.
• Yes. Someone will read, file, print or forward the translated pages: use a tool that produces a rebuilt PDF, and keep ChatGPT for follow-up questions about the content.

The two aren't really competitors. One is a reading tool, the other is a document tool. Most of the frustration people report with ChatGPT and PDFs comes from asking a reading tool to do a document job.