AI PDF Translator: What “AI” Actually Means, and Which Kind You Need

Search for an AI PDF translator and every result claims the label, from a chat window that reads your upload to established translation services that have added machine translation to their workflow. The label is true in all of them and useful in none, because “AI” now covers three genuinely different architectures that produce three different kinds of output. Choosing the wrong one costs you either accuracy or an afternoon rebuilding a document by hand.

This guide is a field guide to what's actually under the hood. It is opinionated about scanned documents in particular — the case where the differences stop being academic — and honest about where each approach is the right call.

Three things that all get called an AI translator

1. Chat assistants (ChatGPT, Gemini, Claude)

A large language model with vision. You upload the PDF, the model reads the page images, and it answers in the chat: translated text, a summary, an answer to a question. Translation quality on the text itself is excellent — these models are the best translators of nuance and idiom currently available. By default, though, the output is a conversation. Ask ChatGPT to keep the layout and it will build a PDF; in our side-by-side test that page was tidy on plain white backgrounds, but photo backgrounds came back as grey blocks and some symbols turned into question marks. Gemini only answers in the chat. There, tables come back as markdown at best, page structure is gone, long files hit upload and response limits, and on multi-page documents the model may compress a section without telling you. We wrote up the specifics in Can ChatGPT translate a scanned PDF?.

2. Neural machine translation with a PDF wrapper

Google Translate's document mode, DeepL's file translation and most “translate your PDF online” sites. A neural MT engine translates the text, and a wrapper tries to put it back into a document. This works well on born-digital PDFs with a real text layer. Feed it a scan and results vary. Some refuse it outright. Many run a generic OCR pass and translate whatever came out, with the tables, columns and stamps of the original reflowed into something that looks like a Word file after a bad paste. The better ones, DeepL among them, rebuild the page reasonably well but still slip on photo backgrounds and figures.

3. Layout-preserving pipelines

A newer category built specifically for scanned and image-only PDFs. The pipeline OCRs the page as an image, decides what is text and what is not (stamps, signatures, rules, figures), translates the text with an LLM, erases only the source text from the page image, and re-typesets the translation into the exact space it vacated. What comes out is a PDF that looks like the original with the language swapped — a lab table still a lab table, a hanko still sitting on the characters it was stamped over. Reglyph is one of these; the translate scanned PDF without losing formatting guide walks through what the reconstruction step actually does.

The question that picks the category

Ignore the feature lists and ask one thing: what do you need to exist when this is done?

• You need to know what a document says → a chat assistant. Fastest, best at nuance, free to try. Stop reading here if this is you.
• You need a translated file of a born-digital PDF and don't care that it's reflowed → an MT wrapper. Cheap, instant, fine for reading copies of clean PDFs.
• You need a translated document that mirrors the original — for filing, for review against the source, for anything with tables, forms, stamps or numbers whose meaning depends on position → a layout-preserving pipeline. The only category where a scanned certificate or a 60-page bank statement comes back verifiable.

What to test before trusting any of them

Marketing copy converges; behavior on a hard page does not. Run one real document through the free tier and check four things:

1. 1Every page comes back. Count them. Silent page loss is the most expensive failure mode in the chat category and it never announces itself.
2. 2The numbers are untouched. Amounts, dates, ID numbers, reference ranges. A translator that “helpfully” reformats a date or rounds a value is disqualified for anything official.
3. 3Tables survive as tables. Open the output next to the original. If a value moved columns, the document is wrong in a way a reader won't notice.
4. 4The stamps are still there. Seals, signatures and handwritten notes should still be there, in place. If they vanished, the tool treated the page as text and threw away everything that wasn't.

Where the categories are heading

The lines are blurring in one direction only: the layout-preserving pipelines are adopting the same frontier language models the chat assistants use for the translation step, so the quality gap on the words themselves is closing to zero. What isn't closing is the output gap. A chat window returns prose unless you ask for a file, and even then the page it builds holds up on plain layouts and slips on busy ones; an MT wrapper will keep reflowing scans because it never looked at the page as an image. If your documents are scans, the choice was made for you the moment they were scanned.

For document-specific requirements — what an insurer, a consulate or a university actually wants back — the document translation hub has a page per document type with the layout details that matter for each. Reglyph's own free tier — 5 pages a day, scanned PDFs included — is described on the free PDF translator page.