Translating Japanese: 10 Pitfalls That Quietly Change the Meaning

Japanese is the language where a translation can be fluent, grammatical, confident, and wrong, with nothing in the output to tell you which. That is what makes it different from French or Spanish. When a translation from those goes bad, it usually goes bad loudly. When a translation from Japanese goes bad, it often reads beautifully.

Here is the shape of the problem, from a translator describing it on Hacker News:

Dalewyn on Hacker News, in a thread on how manga was translated for America.

Two clearly different Japanese phrases collapse into one English sentence. No error is raised. The translation looks complete. Information is simply gone.

This post walks through ten failure modes that show up when you translate into or out of Japanese. Each one gets a real example, an explanation of why it breaks, and a check you can run on a finished translation. If you are new to reading Japanese in the Latin alphabet, the romanizations below follow Hepburn.

Grammar: what the Japanese source does not say #

Three of the most common failures have nothing to do with vocabulary. They happen because Japanese grammar leaves out information that English grammar requires, and something has to fill the gap.

1. The subject is missing, and the translator guesses #

Japanese sentences routinely have no subject and no pronoun. The speaker assumes context supplies it. Take two words:

好きです。(suki desu)

That is a complete, natural Japanese sentence. It can mean "I like it," "I like you," or "she likes him," depending entirely on what came before. Nothing in the sentence narrows it down.

A machine translator has to output something, so it picks. It almost always picks "I like it," the most frequent reading in isolation. When the surrounding paragraph was about a third person, the output is now a confident, grammatical lie.

This gets worse the shorter the text is. One Hacker News commenter put the rule plainly: "the smaller the chunks of texts are the worse a computer will do". UI strings, button labels, chat messages, and subtitle lines are exactly the short, context-free chunks where Japanese leaves the most out.

The check: for every sentence in the English output that contains "I," "you," "he," "she," or "they," find the pronoun in the Japanese source. If it is not there, the translator inferred it. Confirm the inference against the paragraph around it.

2. Number is optional, and counters are not #

Japanese nouns do not normally mark singular or plural. 猫がいます (neko ga imasu) is "there is a cat" and "there are cats" at the same time. Plural markers exist (たち, ら, and reduplication as in 人々) but they are optional and mostly limited to people. English forces a choice on every single noun, so every noun is a small guess.

Going the other direction, English forces a different problem. Japanese cannot say "three files" without attaching a counter, and the specific counter encodes the physical nature of the thing being counted:

Note that 本 is both the word for "book" and the counter for pen-shaped objects, and that the reading changes with the number: 三本 is san-bon, not san-hon. "Three files" has no correct Japanese form until you know whether a file is a sheet of paper, a bound document, or an icon on a screen. There are generic fallbacks (三つ, 三個), but they read as vague in product copy. A translator working from a spreadsheet of UI strings does not know.

The check: search the Japanese output for every number. For each one, confirm the counter matches what the thing actually is in your product, not what the translator assumed it was.

3. The verb comes last, and so does the negation #

Japanese is subject-object-verb, and the verb carries tense, politeness, and negation. All of it lands at the very end of the sentence:

彼は明日の会議に出席します。(kare wa ashita no kaigi ni shusseki shimasu), He will attend tomorrow's meeting.

彼は明日の会議に出席しません。(kare wa ashita no kaigi ni shusseki shimasen), He will not attend tomorrow's meeting.

The two sentences are identical until the final syllable. This is fine for a human reading a whole sentence. It breaks anything that processes Japanese in fragments: subtitle lines split across two frames, a UI string concatenated from two variables, a chat interface that streams a partial sentence, a translator paid per line who is working line by line.

It also means the English rendering has to restructure the whole sentence rather than translate through it. There is no way to start the English sentence until you have read the Japanese one to the end.

The check: never translate a Japanese sentence that has been split across two strings, two subtitle lines, or two table cells. Reassemble it first, then translate.

Register: the decisions Japanese forces you to make #

English lets you write a sentence without deciding how you feel about the person you are writing to. Japanese does not. Every verb you choose encodes a social relationship, and there is no neutral option.

4. Politeness is a choice on every verb #

One Hacker News commenter described Japanese speakers as effectively speaking four languages: "casual speech, normal polite desu/masu speech, humble speech, and keigo". That last pair is not quite a hierarchy: humble and respectful language are two directions rather than two rungs. But the count is right. The same request, "please check this," has at minimum these forms:

An English source sentence carries none of that information. The translator has to supply it, and a machine supplies whatever its training data averaged out to. One commenter comparing the two on exactly this point wrote that "often times the tone from gpt4 are little bit 'rude'/direct or 'weird' in en to jp, deepl is as good as an example sentence they would put into a textbook".

Textbook polite is not always right either. Business Japanese is difficult enough that, as another commenter observed, "even Japanese themselves have to be instructed in how to use it when they enter corporate jobs". If you are sending a Japanese business email, the register is not a stylistic preference. It is the difference between a message that gets answered and one that gets forwarded to someone's manager.

This is the same failure mode that turns a translated cover letter into a desk reject, except that Japanese has four settings instead of two, and they are not a single scale.

The check: decide the register before you translate, not after. Then read the Japanese output and confirm that every sentence sits at the same level. Register that drifts mid-message is the single clearest sign of machine output.

5. There is no neutral word for "you" #

English "you" is invisible. Japanese あなた (anata) is not. Depending on context it reads as distant, textbook-like, confrontational, or as what a wife calls her husband. It is a common default in English-to-Japanese machine output, and it is almost never what a Japanese writer would have used.

Japanese handles the second person in three other ways, in rough order of preference:

1. Omit it entirely. お時間ありますか (o-jikan arimasu ka), literally "is there time," is the normal way to ask "do you have time?"
2. Use the person's name plus さん. 田中さんはどう思いますか (Tanaka-san wa dō omoimasu ka), "what do you think, Tanaka-san," said directly to Tanaka.
3. Use their title. 部長 (buchō, department head) instead of a pronoun.

Marketing copy is where this hurts most. English landing pages are built on "you." A direct rendering produces a Japanese page full of あなた that reads like a badly dubbed advertisement.

The check: count the occurrences of あなた in your Japanese output. In most business or marketing text, the correct number is zero or close to it.

6. Honorifics do not survive the trip #

さん (-san), 様 (-sama), 先生 (-sensei), 先輩 (-senpai), 君 (-kun), ちゃん (-chan). Each one carries information English has no slot for.

田中さん (Tanaka-san) is the clearest case. It is gender-neutral. Every English option, "Mr. Tanaka" or "Ms. Tanaka," adds a fact the Japanese did not state. "Tanaka" alone drops the politeness and, in an English business context, reads as curt. Keeping "Tanaka-san" preserves everything but marks the text as a translation.

先輩 (senpai) is worse, because the English word does not exist. It means someone who entered the same school or company before you, and it carries an ongoing obligation in both directions. "Senior colleague" is the dictionary answer and it is flat. "Mentor" imports a relationship that may not be there.

There is no correct answer here. There is only a decision, and the decision has to be the same on every page.

The check: pick one policy per project (keep honorifics, drop them, or convert them) and grep for exceptions. Mixed policies within one document are the actual defect, not the policy you picked.

7. Speech style encodes gender, age, and role #

Japanese fiction and dialogue use 役割語 (yakuwarigo), role language: the way a character speaks tells you who they are before they say anything about themselves. The first-person pronoun alone does a lot of work. 俺 (ore) is blunt and masculine, 僕 (boku) is softer and male, often younger or more reserved, 私 (watashi) is neutral-polite, わたくし (watakushi) is formal. Sentence-final particles and endings like ぜ, わ, and のよ layer gender and attitude on top, at least in standard Tokyo speech.

English has none of this, which means translating English dialogue into Japanese requires inventing it. Get it wrong and the character is miscast. The site Legends of Localization documents the result: a male character in Stardew Valley who switches between feminine and masculine speech, a demon who speaks in a feminine register, a Harry Potter character who sounds female in Japanese while female characters sound male.

None of this is a translation error in the usual sense. Every individual line is accurate. The character is just wrong.

The check: if you are localizing anything with a voice, a character, or even a brand persona, write a one-paragraph brief per speaker before translation starts: pronoun, politeness level, age, and relationship to the listener. Translators cannot infer this from an English string table, and they will guess differently on different days.