不受歡迎的觀點：到 2027 年，模型將是文件翻譯中最不重要的一環

TL;DR

Most conversations about AI document translation in 2027 still circle one question: which model translates best? I think that's the wrong question. Translation quality at the sentence level is mostly good enough for many language pairs. What breaks in production is everything around the sentence: file structure, context splitting, terminology consistency, and the fact that nobody can verify a fluent 60-page output quickly. If your pipeline is "extract text, send to LLM, paste back," the model is not your bottleneck. Your pipeline is.

Change my mind in the comments.

The comfortable consensus

The story goes like this. LLMs got good at translation. Context windows got huge. So translating a PDF, a DOCX contract or a slide deck is now a solved problem: upload, wait, download.

I've watched enough real documents go through these pipelines to disagree. Not because the models are bad. Because a document is not a string, and most tooling still treats it like one.

1. Documents are not text

A DOCX file is a zipped bundle of XML where one visible sentence can be split across five runs because someone bolded a word in 2019. A PDF is not even a document in the semantic sense. It's a set of drawing instructions that place glyphs at coordinates. Reading order, table cells, footnotes, headers and text boxes all have to be reconstructed before a single word gets translated.

Then the translation has to go back in. German runs about 30% longer than English. Arabic and Hebrew flip direction. Japanese has no spaces to wrap on. Your carefully reconstructed table now overflows, the footnote markers point to the wrong notes, and a merged cell ate half a row.

None of that is a model problem. It's a parsing and layout problem, and it's where most "AI document translation" tools quietly fail.

2. Bigger context windows didn't fix chunking

Yes, you can fit a long contract in one prompt now. In practice, most pipelines still segment documents for cost, latency, retries and layout mapping. The moment you segment, you lose document-level coherence:

• "Agreement" on page 2 becomes "Contract" on page 14.
• A defined term in the glossary section gets translated three different ways.
• A pronoun that depends on a paragraph two chunks back gets the wrong gender.

Each chunk reads beautifully on its own. The document as a whole is inconsistent, and in legal, medical or technical content, inconsistency is an error.

3. Fluent and wrong is worse than clunky and wrong

Old statistical MT was obviously bad when it was bad. You could see the seams. LLM output is fluent even when it omits a clause, drops a "not," rounds a number, or confidently invents a smoother phrasing that changes the meaning.

This flips the risk model. The danger is no longer text that looks broken. It's text that looks perfect and isn't. A reviewer skimming a fluent translation is far less likely to catch a missing negation than one wading through awkward phrasing.

4. "Which model is best" doesn't have a stable answer

Leaderboards are averages. Your document is not an average. The model that wins on English to Spanish marketing copy can lose on English to Thai pharma labels or Polish to English legal terms. Rankings also shift every few months as new versions ship.

So instead of picking a champion, I've become more interested in approaches that run the same document through several engines and look at where they disagree. Disagreement is a cheap, surprisingly useful signal for "a human should look here." MachineTranslation.com's document translation is one example of this compare-and-agree approach applied to full files, and it's a useful reference point if you want to see how it behaves on your own documents.

To be fair to the counterargument: agreement is not proof of correctness. If every model shares the same blind spot (a rare term, a regional legal concept), they can all agree on the same wrong answer. Consensus narrows where you need to look. It doesn't remove the need to look.

5. Human review isn't disappearing. It's changing shape

The "AI will replace translators" take and the "AI translation can't be trusted" take are both lazy. What I see happening is a role shift: from translating every sentence to verifying the risky ones. The skill that matters in 2027 is triage. Knowing which 5% of a document carries 95% of the risk (numbers, names, obligations, dosages, negations) and spending human attention there.

What I'd actually test in a document translation pipeline

If you're building or choosing a tool, here's what I'd check before arguing about models:

1. *Round-trip formatting: * Translate a messy real DOCX and PDF (tables, footnotes, tracked changes, RTL text). Open the output. Does it still look like the original?
2. Entity and number diffing: Automatically compare every number, date, currency, URL and proper noun between source and target. This catches a shocking amount.
3. Terminology consistency: Pick 10 defined terms and count how many distinct translations each one gets across the file.
4. Omission detection: Compare segment lengths and sentence counts. Big gaps are a red flag for dropped content.
5. Disagreement mapping: Where do two or more engines diverge? Route those segments to review first.
6. **Skip back-translation as your main QA. **It's popular, but a model translating its own output back often smooths over the exact errors you're trying to find.

Where I might be wrong

• Maybe native multimodal models will read PDFs as images and regenerate layout well enough that parsing stops mattering.
• Maybe one model will pull so far ahead on every pair that ensembles become unnecessary overhead.
• Maybe review tooling will get good enough that "fluent and wrong" is caught automatically.

I don't think any of those land in 2027. But I'd genuinely like to hear from people shipping this stuff.

Quick FAQ

What is AI document translation? AI document translation is the process of translating entire files (PDF, DOCX, PPTX, XLSX) with machine translation or large language models while preserving the original formatting, layout and structure.

**Is AI document translation accurate enough in 2027? **At the sentence level, often yes for common language pairs. At the document level, accuracy depends on parsing, chunking, terminology consistency and verification, not only on the model.

What is the hardest part of translating a PDF with AI? Reconstructing the reading order, tables and layout before translation, and fitting translated text back into the original design afterward.

Do you still need human review for AI-translated documents? For anything with legal, medical, financial or safety consequences, yes. The efficient approach is targeted review of high-risk segments rather than line-by-line retranslation.

How can I check AI translation quality without speaking the target language? Diff numbers and named entities, check terminology consistency, compare outputs from multiple engines to find disagreements, and send flagged segments to a qualified reviewer.

Over to you

• What's the worst document translation failure you've seen in production?
• Is your bottleneck the model, the parser, or the review step?
• Single best model, or multiple engines cross-checked: which camp are you in, and why?

Drop it in the comments. I'll reply to everyone.