October 28, 2026
Convert your text to a supported file, upload the whole thing instead of pasting it in sections, pin the names and terms that must not change, and read the consensus signals before you read the translation. A book, a research paper, or a long article goes through MachineTranslation.com as one file up to 70MB, and the places where the models disagree tell you where to slow down.
A long text can fail in places a single paragraph never does. The file can be in a format the translator cannot open. The length can run into a limit. And over hundreds of pages, a name or a term can come back spelled differently on different pages, with nothing to tell you it happened. We built SMART, the consensus system behind MachineTranslation.com, to put a signal on exactly that kind of silent failure.
This guide covers translating for your own reading. If you plan to publish or share a translation, get the rights holder's permission first. If you will cite what you read, see our guide for academics on reading versus citing.
Yes. Upload the book as one file and you get one translated file back. What no AI model can give you is a guarantee that every sentence is right, and a book has tens of thousands of them.
A single model can invent a detail, flatten an idiom, or drop a term, and nothing flags it, especially when you do not read the source language. MachineTranslation.com runs the same text through up to 22 independent models and surfaces the version most of them agree on. When the models converge, that agreement is the quality signal. When they diverge, the divergence marks the passage for a closer look.
We tested what divergence looks like on a short service agreement translated into Filipino. Eight models reached 85% agreement and disagreed on 5 terms. One of the five was the title: the models split between two valid renderings of “Service Agreement” with different formal registers, and a single model would have handed you one of them with no sign that the other existed. Stretch that across 300 pages and a chapter title, a place name, or a character's form of address can wobble the same way.

Expect the weakest results on poetry, wordplay, dialect, and humor, where meaning sits in the sound or the cultural reference. Read those passages with the original beside you.
The translation step is the same for all three. The file you start with and the failures you watch for are not.
| Text | Typical file | What tends to go wrong | What to do |
|---|---|---|---|
| Book | EPUB, Kindle, or a scanned PDF | The format is not supported. Names and tone drift over many chapters. Wordplay and idiom | Convert a DRM-free copy to DOCX, PDF, or TXT. Pin character and place names. Read idioms against the original |
| Research paper | A PDF, often two columns with tables | Layout, abbreviations, technical terms, numbers and units, equations saved as images | Upload the PDF itself. Pin the technical terms. Check numbers and units against the original |
| Long article | A web page or a PDF printout | Pasted text loses its structure and hits the 3,000-word text box limit. Quotes and names | Save the page as a PDF or HTML file and upload it. For HTML, only visible text is translated and tags stay intact |
For papers, the layout question is the one we tested. A one-page, two-column PDF translated from English to German by eight models kept both columns, a table, and a callout box in place. A title block, a bullet list, and a word in bold slipped. Our test file had no equations or figures, so we make no claim about them.

Five numbers decide this, and they come from the pricing page and the Document translation page. The example is a typical novel of about 90,000 words.
| Limit | The number | What it means for the novel |
|---|---|---|
| File size | 70MB per file | A text-only book is far under it. Our 54,000-word sample handbook is about 32KB as a Word file. Scans are heavier, because every page is an image |
| Text box | 3,000 words per click | The novel would take 30 separate pastes, with nothing tying them together |
| Free-plan preview and unlock | A partial free preview of long text, then $2.50 per 300 words to unlock | The novel is 300 blocks of 300 words, or $750 at the unlock rate, so price that against a 1-day plan or Pro |
| Layout downloads | A fair-use limit of 100,000 words a day | The novel fits in one day. A 150,000-word book needs two |
| Human Verification | Free for up to 2,000 words a month on paid plans | About 2% of the novel, so keep it for the passages that carry the story or the argument |
Five steps, in the order that keeps a long translation consistent.
MachineTranslation.com accepts PDF, DOCX, XLSX, PPTX, TXT, CSV, SRT, HTML, XML, JSON, RTF, ODT, and image files. EPUB and Kindle formats are not on that list, so convert a DRM-free ebook to DOCX, PDF, or TXT with an ebook converter. A copy locked by DRM generally will not open outside its reading app, so start from one you are allowed to use. A scanned book is read with OCR before translating and comes back as an editable Word file.

Drag the file into the upload box on the Document translation page. On the Free plan or with no account, long text may show a partial preview first. The full translation is unlocked with a one-time payment, or by moving to Pro or a 1/3/5-day plan, which unlock translations automatically. Decide this before you upload, using the limits table above.

When the file is ready, open the consensus view. It shows how many models worked on the file, the percentage in agreement, and how many terms they disagreed on. We treat 90% agreement with zero or one disputed term as a strong candidate to read as it is, and agreement in the mid-80s or several disputed terms as a reason to read more closely.
That rule comes from everyday business text. For a novel or a paper, use it as a map of where to slow down.
For very long documents, MachineTranslation.com shows the source and the translation side by side, segment by segment, so you can check a line without losing your place in the whole. Use it where your understanding depends on one word: a clue in a mystery, a number in a paper, a quoted line you plan to repeat.
If a passage reads stiffly, the Translation team panel names the model that read most natural for your text, and you can open that model's tab to compare the sentence. I would start with the first chapter and the passages where the models disagree.

The translation is for reading. If you will cite a paper, cite the original. If you want to share a translation, get the rights holder's permission first.
Consistency is the failure you cannot see from one page. In our internal analysis, documents processed in fragments showed a 28% higher rate of terminology inconsistency than documents processed whole, which we describe in our guide to getting better results from AI translation. When you paste a book in sections, each section is translated without knowing how the others were rendered, so terms and phrasing drift.
A single upload lets up to 22 models work on the full document. If a limit does force a split, cut at chapter breaks and reuse the same list of terms for every part.
You cannot proofread 300 pages in a language you do not read, so check by sampling. Run these five checks on a few chapters or sections, and run them again on any passage the consensus view flagged.
A mismatch between two chapters, or between the translation and the original on a passage you care about, is a question to resolve. It does not prove the translation wrong, and it does not prove it right.
With the file converted, the terms pinned, and the agreement signals read, you can follow a long text in a language you do not read and know which passages deserve a second look. You can upload your book, paper, or article on MachineTranslation.com.
Yes. Upload the book as one file and you get one translated file back. No AI model guarantees every sentence, so read the consensus signals first and check the passages where the models disagree against the original.
Upload the PDF itself on MachineTranslation.com, choose your language, and download the translated file. A scanned PDF is read with OCR first and comes back as an editable Word file.
EPUB and Kindle formats are not on the supported list for now. Convert a DRM-free copy to DOCX, PDF, or TXT first. A copy locked by DRM generally will not open outside its reading app.
Yes. In a one-page, two-column test, the columns, a table, and a callout box stayed in place, while a title block, a bullet list, and bold text slipped. That test had no equations or figures, so check those against the original.
It depends on your plan, because the Free plan unlocks long text at $2.50 per 300 words, which is $750 for a 90,000-word novel, while Pro and 1/3/5-day plans unlock translations automatically. Price the plans against the unlock before you upload.
Every tool sets its own file-size and length limits, and a chat window handles a long text in pieces, each translated with one model's judgment. MachineTranslation.com runs ChatGPT as one of up to 22 models, takes the whole file in one upload, and returns the version most of them agree on.
Translating for your own reading is different from publishing or sharing a translation, which needs the rights holder's permission. This is general information, not legal advice.

By Rachelle Garcia
Connect on LinkedInRachelle leads product and AI at Tomedes, where she runs the experiments that turn internal data into better translation experiences. She writes about what actually happens when you build AI products such as MachineTranslation.com — the numbers, the surprises, and the parts that don't go to plan.