Why AI Struggles With Japanese, Chinese, and Korean Translation and How to Compensate

A client sends a 40-page product manual and asks for Japanese, Simplified Chinese, and Korean. You run it through the same workflow that has handled German and Spanish all year. The German comes back clean. The Japanese comes back fluent, confident, and wrong in ways nobody on your team can see. AI translation for Asian languages fails differently than it fails in European pairs, and the difference isn't just one of degree. We've watched enough CJK projects go sideways to know the problems cluster in a few predictable places, and that most of them get solved during preparation rather than during review.

What actually goes wrong with AI translation for Asian languages

Start by separating fluency from accuracy. Modern models write excellent Japanese. The output reads naturally, uses appropriate connectives, and doesn't sound machine-generated to a casual reader. That's exactly what makes the failures hard to catch. In European pairs, a bad machine translation usually reads badly, so reviewers find it. In CJK output, a wrong translation often reads better than a correct one, because the model resolved an ambiguity confidently instead of preserving it.

Underneath that, a few structural problems keep showing up in non-European language pairs.

Japanese and Chinese are written without spaces between words. Every tool in the chain that assumes whitespace tokenization behaves oddly on them, from word counts to QA checks to fuzzy matching against a TM.

All three languages omit information that English forces you to state. Chinese verbs don't inflect for tense. Japanese and Korean drop subjects freely. None of the three mark plurality the way English does. Translating into these languages means throwing information away. Translating out of them means inventing it.

And politeness is grammatical, not stylistic. In Japanese and Korean, register is baked into verb endings rather than word choice, and there's no neutral option to fall back on. The model picks one. If you didn't tell it which, it picks based on its training distribution rather than your client's brand voice.

None of this is exotic linguistics. It's the working reality of any project pairing English with a CJK target, and it changes what your preparation step has to do.

Japanese: the dropped subject and the politeness level nobody chose

Japanese sentences routinely omit the subject when context makes it recoverable. A safety manual that says "must be inspected before each shift" translates cleanly into Japanese without ever naming who does the inspecting. Going the other direction, the model has to supply a subject the source never contained, and it will guess. We've seen an operator's instruction turned into a statement about the equipment itself, which reversed the responsibility described in the original.

The politeness problem is more common and much easier to prevent. Japanese verb endings encode register: plain form, polite form (です・ます), and humble or honorific forms above that. A product manual, a marketing email, and an internal HR policy sit at different points on that scale. If your prompt says nothing about register, the model defaults to です・ます for most content. That's usually acceptable for documentation and usually wrong for anything the client thinks of as brand voice.

A concrete case from a technical documentation project. The source was a set of installation warnings written in imperative English: "Do not connect the unit before grounding." The first pass came back in polite form throughout, which softened directives that the client's Japanese style guide required in plain imperative (連結しないこと). The translation was accurate. It was also unusable without a full pass to change every verb ending. Re-running with one line in the prompt specifying plain imperative form for warning text fixed it in a single run.

So treat register as a preparation decision, not a review finding. Two sentences in the translation prompt naming the target register, plus a note on whether the reader is addressed directly, prevent most of that rework. Worth adding a line on whether the client wants katakana loanwords or native equivalents for technical terms, since models lean heavily on katakana when left alone.

Chinese: measure words, script variants, and grammar that isn't there

Simplified and Traditional Chinese aren't a formatting toggle. The vocabulary differs, the terminology conventions differ, and mainland, Taiwan, and Hong Kong readers expect different terms for the same technical concept. A model asked for "Chinese" produces Simplified by default and may mix conventions if the prompt is vague. Specifying zh-CN, zh-TW, or zh-HK is the minimum, and for terminology-heavy content the glossary needs to be variant-specific rather than shared across all Chinese targets.

Measure words cause a quieter class of error. Chinese requires a classifier between a number and a noun, and the correct classifier depends on the noun's shape or category. Models get the common ones right and improvise on domain-specific nouns. In a parts catalogue we reviewed, several component types were counted with 个, the general-purpose classifier, where the client's existing documentation used more specific ones. Nothing was incomprehensible. It read as if written by someone competent but outside the industry, which was precisely the impression the client didn't want.

The absence of tense and plural marking cuts both ways. Translating English into Chinese, the model drops that information, and it should. Translating Chinese into English, the model has to reconstruct it, and this is where hallucination-adjacent errors appear. A specification written in Chinese without tense marking becomes an English sentence in the past perfect, implying a completed action the source never asserted. On regulatory and contractual content that matters enough that we treat Chinese-to-English as a different risk category than English-to-Chinese.

One practical mitigation: give the model document-level context rather than isolated segments. Tense and number are usually recoverable from surrounding sentences, and a model that can see the whole paragraph reconstructs them far more reliably than one working segment by segment.

Punctuation is a smaller issue that clients notice anyway. Chinese uses full-width punctuation, including its own comma, period, and enumeration comma (、). Models mix half-width and full-width marks when the source is English, especially around numbers and parenthetical text. A reviewer who reads no Chinese can still spot a Latin comma sitting inside a run of Han characters, so this belongs on your automated check list rather than in a linguist's hours.

Korean: speech levels and the spacing nobody checks

Korean has spaces, which makes it look more tractable than Japanese or Chinese. The spacing rules (띄어쓰기) are genuinely hard, though, and native speakers disagree about edge cases. Models produce spacing errors that are invisible to a non-Korean reviewer and immediately noticeable to a Korean reader. They read as sloppy rather than as wrong, which is arguably worse for a client deciding whether to keep working with you.

Speech levels work like Japanese register but with more distinctions in active use. The formal polite ending (합니다체) suits documentation and official notices. The informal polite ending (해요체) suits customer-facing marketing and app interfaces. Mixing them inside one document is the single most common complaint we hear from Korean reviewers looking at AI output, and it happens because the model treats register as a local decision per segment rather than a property of the whole document.

Honorifics compound this. Korean marks respect toward the subject of a sentence through verb infixes and separate noun forms, and the right choice depends on who's reading and who's being discussed. An HR policy referring to employees, managers, and executives may need different honorific treatment for each. No amount of general instruction covers it, so this is a case where we tell agencies to plan on a Korean reviewer rather than treat one as an optional check.

Another example worth recording. A software company translated its onboarding emails into Korean with a well-built glossary and got terminology right throughout. The emails still had to be redone, because the model alternated between 합니다체 and 해요체 across paragraphs. The glossary controlled nouns and controlled nothing about verb endings. Register belongs in the prompt, stated as a document-wide rule.

Why a glossary behaves differently in CJK

A glossary in a European pair mostly works by substitution. You list "circuit breaker" against its target, and the model uses that term. In CJK targets, the listed term often has to change shape depending on where it lands in the sentence, and a rigid instruction to use the exact string can produce grammatically broken output.

Japanese attaches particles to nouns and conjugates verbs pulled from a glossary. Korean adds postpositions whose form depends on whether the preceding syllable ends in a consonant. Chinese is more forgiving, though compound terms still get split or reordered in ways a simple string match won't catch. Build glossaries for CJK the way you build them for French and your QA check will report violations that aren't violations, while the real terminology drift hides inside the noise.

A tiered approach works better. Brand names, product names, model numbers, and legally fixed terms go in a strict do-not-translate or must-match list, since those genuinely should appear verbatim. Ordinary domain terminology goes in a softer list described as preferred terminology, and the model gets to inflect it. Our guide on how to choose which terms go in your AI translation glossary covers the tiering logic in more depth, and the reasoning applies with more force in CJK than anywhere else.

The other adjustment is script-level. A Japanese glossary entry may need both the kanji form and the katakana form, because the source document might use either. A Chinese glossary needs separate entries per variant. Small additions, and they turn a glossary from a source of false QA flags into something that actually constrains the output.

How to review output in a language nobody on your team reads

Most agencies working into CJK don't have in-house Japanese, Chinese, and Korean reviewers. The practical question is what you can verify yourself and what you have to pay a linguist for.

Non-linguistic checks travel across scripts and are worth automating. Numbers, dates, currency symbols, model numbers, URLs, and units either match the source or they don't, and a machine comparison catches discrepancies regardless of script. Untranslated segments are next: CJK output containing runs of Latin characters is either correct (a brand name, a code sample) or a segment the model skipped. Both cases are findable without reading the language.

Completeness is the third thing you can check. Because CJK text is denser than English, a translated Japanese paragraph is visibly shorter than its source, and length ratios that look alarming in a European pair are normal here. What isn't normal is a segment that came back empty or drastically truncated. Set your expectations by measuring a few known-good documents rather than by applying European length heuristics.

Past that you need a native reviewer, and the useful move is to scope their time rather than skip them. Send a bilingual source-and-target spreadsheet rather than the finished document, ask specifically about register consistency and terminology, and hand over the glossary and prompt you used. A reviewer who knows the intended register can check 40 pages for register violations far faster than they can do a general quality read. Our post on how to post-edit AI translations efficiently works through the scoping logic in general terms.

This works best when the client has a style guide or previously approved material to anchor the review. It doesn't apply to legally binding content, where full human translation remains the defensible choice no matter how good the AI output looks.

What to change in your next CJK project

Do one thing differently before the next Japanese, Chinese, or Korean job: write the register decision into the prompt before you run anything.

Name the target speech level or politeness form as a document-wide rule (plain, polite, or formal for Japanese; 합니다체 or 해요체 for Korean), specify the exact locale for Chinese, and say whether the reader is addressed directly. Three sentences. Then run a short sample, two or three pages, and have a native speaker confirm the register before you commit the full document. That sample review costs an hour and prevents the situation where a 40-page translation is accurate and still has to be redone from the first verb ending onward.

Terminology is the part most teams already handle. Register is the part that gets discovered at delivery.