The evidence
In the 2025 WMT evaluation — 30 language pairs, 60 systems, professional annotators marking real errors — the human translator scored 89.2 into Japanese. The best machine system, Gemini 2.5 Pro, scored 85.8.
The human still wins. That is no longer true of most European pairs, and it is worth understanding why it remains true here.
Why Japanese is hard for machines
The information is missing from the source. Japanese requires the speaker to encode their relationship to the listener — who is senior, how formal the situation is, how much distance to keep. English does not carry that information anywhere in the sentence. So a system translating English into Japanese must infer something the source text never states, from context it usually does not have.
A peer-reviewed 2026 study tested this on business emails. A plain "translate this" instruction produced "only limited adaptation toward Japanese language patterns", failing specifically on honorifics, formal address, indirectness in hierarchical contexts, and appropriate directness between a superior and a subordinate.
The study's finding about the fix is the commercially interesting part: what corrected the output was supplying context in the instruction. Which is to say the fix was human expertise about the situation, not a better engine. That is exactly the work that is billable, and exactly the work that does not disappear when the next model ships.
Named entities are a coin flip. A 2025 benchmark testing translation of names, brands and cultural references found the widest performance spread of all languages tested in Japanese and Korean. Product names, company names, transliteration conventions, honorific suffixes — the places where a Japanese reader immediately knows whether the text was written by someone who understands the market.
Real commercial content scores worse than benchmarks suggest. On one published scoreboard of thousands of evaluations across real client projects — software, marketing and product content rather than general documents — Japanese scored among the lowest of ten languages tested. Treat that source with more caution than the academic benchmark, but the direction is consistent with what reviewers see in practice: benchmark documents are easier than the content companies actually ship.
What we do
- We fix the register decision at the start. Who is speaking, to whom, in what relationship. It cannot be inferred reliably from the source and it should not be left to the engine.
- We check honorifics and speech level as a named error category, scored separately from accuracy, because output can be accurate and still be socially wrong in a way that costs a relationship.
- We hold entity and terminology decisions across a document set, so product names and transliterations do not drift between files.
- We review layout consequences. Line breaking, vertical text where it applies, mixed-script rendering, and the effect of Japanese text expansion or contraction on a fixed layout.
Where Japanese work is worth most
Japanese crossed with a technical or regulated domain: medical device documentation, automotive service and diagnostic material, patents, and precision manufacturing. Japan remains one of the largest markets in the world for exactly these content types, and the tolerance for translation that reads as foreign is low.
Published 2026 rate data puts English↔Japanese at the top of the major pairs, roughly one and a half to two times English↔Spanish, with specialist work above that. The constraint cited is the supply of qualified translators, and that supply is thinning rather than growing.