← Blog · 2026-06-23

Why machine translation breaks on technical documents

Author: Yuri Vodostoy — Technical translator and founder of Insight Translation

Modern machine translation is genuinely good at the level of a single sentence. The failures show up over a whole document, which is exactly where technical, medical and legal work lives. Three things go wrong, and none of them are about grammar.

1. Terminology drifts

MT translates sentence by sentence with no memory of what it chose two paragraphs ago. The same source term comes out differently across a document. A real example — German → English, the word Leistung in two adjacent sentences of a pump spec:

One term, two translations from the same engine, and the second one is wrong. Now the reader has to guess whether "output" and "performance" mean the same thing. HardTrans reads the whole document first and locks one term to one translation.

2. Register is wrong

Normative documents express requirements with shall, not must. MT defaults to "must" because it's the literal modal — fine in prose, wrong in a specification. Across a requirements document that single substitution appears hundreds of times and quietly changes the legal weight of every clause.

3. Designations aren't resolved

A material number like 1.4541 is left as a bare string. A specialist (or a glossary) renders it as X6CrNiTi18-10 (AISI 321) — the designation the target-language reader actually uses. MT has no glossary, so it can't.

The fix isn't a better sentence model

It's reading the whole document, holding terminology constant, treating your glossary as an invariant, and following the register of the domain. That's the difference between an engine that does fine on a tweet and one you can point at a 200-page spec.

The example above is real output: Google Translate vs HardTrans (Claude Sonnet 4.6 with a client glossary), June 2026.