What the evidence says
English to Korean was one of only two pairs in the 2025 evaluation scored under the stricter MQM protocol, where annotators count and weight actual errors. The human translation tied for first place. No machine system beat it.
The second finding is the more commercially useful one. On real client content, Korean shows one of the widest gaps between the best engine and a reasonable second choice of any language measured — 7.5 points. Korean is not a language where you can pick an engine by reputation and move on.
Why it happens
Korean, like Japanese, encodes the speaker–listener relationship in the grammar, through a system of speech levels that the source English never states. The machine has to infer it. It also has to handle honorific verb forms, address terms tied to seniority, and a register distinction between spoken and written Korean that has no English equivalent.
Korean is not a low-resource language. There is plenty of Korean text. The difficulty is that the information the translation needs is not in the source at all.
What it means for your workflow
Two things. Your reviewers need to be checking speech level and address, which requires a native speaker and a brief — not a glossary. And if you standardised your engine choice on your European results, you have very likely made the wrong choice for Korean and are paying for it in post-editing effort you have never attributed to the engine.
What we do
We review for speech level, address and register consistency. And because engine choice matters this much here, we will benchmark your current engine against alternatives on your own content, and give you the numbers. We do not sell an engine, so we have no answer to defend.