German

The sentences are fine. The document is not.

--:--:-- [PST]
Not evaluated
English to German was not put through human evaluation in 2025 — it sat in the automatic subtrack
WMT25
90.7 vs 82.8
best system against the human reference on Czech to German — machines are strong here
WMT25
A different engine wins
German is one of the few languages the overall best-performing engine does not win
Engine index, 2026
S01

What the evidence says

We will be careful here, because the honest position is more useful than an overstated one.

The 2025 international evaluation did not put English to German through human evaluation at all — it sat in the automatically-ranked subtrack. So there is no independent human-versus-machine result for German, and anyone quoting one to you should be asked for the source.

What the evidence does show: on Czech to German, the best system scored 90.7 against a human reference's 82.8 — machines are strong in German. And on real commercial content, German scores 78.2, mid-table, with one notable detail: German is one of the few languages where the overall best-performing engine is not the best engine. A different model wins it.

S02

Why the difficulty is real anyway

German quality problems are not sentence problems. They are document problems.

German forms compounds freely, so a single technical concept becomes one long word — and a system translating segment by segment will render the same concept three ways across a manual, all of them defensible in isolation. German also carries a formality distinction, du and Sie, that English does not mark at all, so the machine infers it, and it can drift mid-document. Sentence structure places the verb at the end of subordinate clauses, which produces long dependencies that fragment when a segment boundary falls in the wrong place.

None of that is visible at the segment level, which is where your automated checking looks.

S03

What it means for your workflow

German is a language where high automation is reasonable and unattended automation is not. Your engine will produce good German sentences. It will not guarantee that page 180 uses the same term as page 4, or that your product stayed on Sie throughout.

S04

What we do

We check German at document scale: terminology consistency across the full deliverable, compound handling against your glossary, and formality register from first page to last. We report the pattern so it can be fixed in your glossary and prompt rather than in every file. And given that German is one of the languages where engine choice diverges from the default, it is worth an hour of benchmarking.

S05

Ask us to check terminology consistency across a full German document set

Not one file. The set — where the drift actually happens.

Send us a file to evaluate