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

Yes — ChatGPT can translate a scanned PDF, and for a short document where you just need to know what it says, it does the job well. Modern vision models genuinely read images of pages, including scans and phone photos, and the translation quality on the text itself is good. The catch is what comes back: translated text in a chat window, not a translated document. If your endpoint is a PDF that looks like the original — same tables, same stamps, same page structure — a chat interface is the wrong shape for the task, and no amount of prompting changes that.

This post walks through both halves honestly: the cases where ChatGPT is genuinely the fastest option, the specific ways it breaks down on real multi-page documents, and what to reach for when you need a document out, not a conversation.

What ChatGPT actually does with a scanned PDF

When you upload a scanned PDF to ChatGPT, there's no selectable text layer for it to extract — the pages are images. So the model reads the page images directly with its vision capability, the same way it reads a photo. This works far better than people expect. It handles ordinary printed scans in most major languages, copes with moderate skew and noise, and can answer questions about the content, summarize it, or translate it on request.

What it produces, though, is prose. Ask it to translate a scanned invoice and you get the translated text back as chat output — sometimes with a markdown table if the model decides to build one, sometimes as running paragraphs that flatten the table entirely. The original page image is never modified. There is no step where the source text gets erased and the translation gets placed back where it was. That reconstruction is on you.

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 — "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, upload the scan, ask for a translation, and you're done in under a minute. It's a genuinely good tool for this, and pretending otherwise would be dishonest.

Where it breaks down

Multi-page documents hit limits

Long scanned PDFs run into two ceilings at once. Upload limits cap how much you can attach in one go, and response length limits cap how much translated text comes back in one reply. A thirty-page scanned contract typically means splitting the file, feeding it in chunks, prompting "continue" repeatedly, and stitching the pieces together afterward — while checking that the seams line up.

Pages get skipped silently

This is the failure mode that actually burns people. On a long document, the model may summarize a stretch of pages instead of translating them, or quietly compress repetitive-looking sections — and nothing in the output flags that it happened. Unless you cross-check page by page against the source, you won't know a page is missing until someone downstream notices. For anything where completeness matters, that's disqualifying without a manual audit.

Tables and forms come back flattened

A customs declaration or financial statement is mostly structure: rows, columns, boxes, numbers that mean something because of where they sit. Chat output linearizes all of that. Even when the model emits a markdown table, column alignment with the original is approximate, merged cells get lost, and marginal text — stamps, footnotes, handwritten annotations — tends to vanish or get folded into the wrong place.

There's no document at the end

The fundamental issue: you asked for a translated PDF and you have a chat transcript. Turning that transcript back into a document means rebuilding the layout by hand in Word or a design tool — recreating every table, repositioning every field. For a one-page letter that's ten minutes. For a forty-page technical manual it's a project.

The document-out alternative

The gap ChatGPT leaves is exactly the part a purpose-built scanned-document translator handles. Reglyph works on the page image itself: OCR reads the text, the original text is erased from the image, and the translation is typeset back in the same position — so tables, stamps, figures, columns, and numbers stay where they were, and what you download is a finished PDF rather than text to reassemble. Every page in the file gets processed; nothing is silently summarized. It runs in the browser, including on a phone, handles 12+ languages, and offers a bilingual side-by-side export when you want the original and translation next to each other. The first 5 pages are free with no credit card, and paid plans start at $5. It's machine translation, so the same honesty applies as with any AI tool: clear printed scans work best, heavy handwriting is unreliable, and official submissions still need a human translator's certification on top.

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 translating PDFs" comes from asking a reading tool to do a document job.