Why Machine Translation Still Struggles With Japanese

Japanese did more damage to a test paragraph than any of the other 19 languages we measured in our damage ranking. It beat Korean, Chinese, Mongolian, and Arabic, and it did so while producing output that reads perfectly well in English.

That combination is the whole problem. Modern machine translation handles Japanese vocabulary and grammar well enough that the failures don't look like failures. They look like fluent English with something quietly missing.

Here is what that looks like in practice, one failure at a time. Every example below is real mixer output, measured on Google Translate in August 2026.

One Cat, Or Possibly Several

Start with a sentence a first-year student could read.

猫が庭にいます。

The machine returns "There's a cat in the garden."

That translation is fine, and it is also an invention. Japanese does not mark plural on nouns. The sentence says there is cat-presence in the garden. Whether that means one cat or six is not encoded anywhere, and Japanese speakers don't feel the absence because context normally settles it. English has no way to stay neutral, so the machine picks the most common option and moves on.

The same thing happens to counters, which are the part of Japanese that carries the information English is missing:

In: 猫を二匹と本を三冊買いました。 Out: I bought two cats and three books.

The Japanese uses 匹 for the cats and 冊 for the books, two different counting words chosen because one thing is a small animal and the other is a bound volume. English has numbers and nothing else. The count survives, the classification is deleted, and there is nowhere to put it.

The Most Common Sentence in Japanese Does Not Survive

This is the result that best captures the problem:

In: 来年もよろしくお願いします。 English: I look forward to working with you again next year. Back to Japanese: 来年もまたご一緒できることを楽しみにしています。

よろしくお願いします is arguably the most-used polite formula in the language. It has no English equivalent because it isn't a statement, it's a social move, somewhere between "please treat me favorably," "thanks in advance," and "let's do this." The English rendering is a reasonable guess at what it accomplishes in this context.

The return trip is where you can see the damage. The machine translated the English back literally and produced a grammatical Japanese sentence that no one would write here. The formula is gone and a description of the formula's effect took its place.

The same pattern hits Japanese apologies, which are far more specific than English ones:

In: 気を遣わせてしまってすみません。 English: I'm sorry for causing you trouble. Back to Japanese: ご迷惑をおかけして申し訳ありません。

The original apologizes for making someone be considerate on your behalf, which is its own social category in Japanese. The return apologizes for inconveniencing them. Related, not the same, and the specific thing being apologized for is gone.

Favors Have Direction, and the Direction Falls Off

Japanese verbs of giving and receiving encode who benefited. English has to say it separately or not at all.

In: 友達の宿題を手伝ってあげた。 English: I helped my friend with their homework. Back to Japanese: 私は友達の宿題を手伝った。

The original ends in あげた, marking that the speaker did the friend a favor. The English has no room for that, so it disappears, and the Japanese that comes back is a flat statement of fact. It also opens with 私は, an explicit "I" that the original didn't need and that real Japanese usually omits.

The reverse construction survives better, which is the useful counterexample:

In: 友達が宿題を手伝ってくれた。 Out: My friend helped me with my homework.

くれた marks a favor coming toward the speaker, and English recovers it as "helped me." When Japanese encodes something explicitly enough, the machine reads it accurately. The failures cluster entirely around what the language leaves to context or expresses through structure English doesn't have.

Elephants Have Long Trunks

Japanese builds sentences around a topic rather than a subject, and the classic demonstration is this one:

In: 象は鼻が長い。 English: Elephants have long trunks. Back to Japanese: 象は長い鼻を持っている。

The original says, roughly, "as for elephants, the nose is long." There are two things marked in that sentence and neither of them is a grammatical subject in the English sense. The machine resolved it into possession, which is the only structure English offers.

Then the return trip carried the possession back into Japanese as 持っている, literally "is holding." Japanese does not describe body parts that way. The sentence is grammatical and slightly wrong in a way that is hard to point at, which is the register this entire post lives in.

Pronouns That Were Never There

Japanese omits subjects across whole stretches of text. Watch the machine invent three of them:

In: 田中さんに会った。元気そうだった。仕事を辞めたらしい。 Out: I met Tanaka-san. He seemed well. Apparently, he quit his job.

Not one of "I," "he," or "he" appears in the Japanese. All three were inferred, and all three are plausible. They are also unverifiable. The second sentence could as easily be about a third person, and nothing in the original settles it.

The return trip also upgraded the politeness, turning the plain 会った into the humble お会いしました, which changes who the speaker is talking to. Nobody asked for that either.

Going the other direction, Japanese sometimes resolves an ambiguity English was happy to leave open:

In: Tanaka told Suzuki that he had made a mistake. Japanese: 田中は鈴木に、自分が間違いを犯したと告げた。 Out: Tanaka told Suzuki that he had made a mistake.

The English "he" could be either man. Japanese has no comfortable way to be vague there, so the machine chose 自分, which points back at the subject and pins the mistake on Tanaka. Then the round trip returned English that is ambiguous again, hiding the fact that a decision was made at all. If you were translating a legal document, that is the sentence that costs you.

The Eel Problem

The classic example linguists use for Japanese context dependence:

私はうなぎです。

The machine translates it as "I am an eel," which is exactly what it says and almost never what it means. In a restaurant, this is how you order. The topic marker は sets up "as for me," and the situation supplies the rest: as for me, it's the eel. Japanese linguists have written papers on this construction. Machine translation reads it as a claim about species.

You will not hit this in a business document. You will hit it constantly in dialogue and subtitles, which is exactly where translated Japanese tends to read like a robot.

The Underlying Problem

English and Japanese disagree about which facts a sentence is required to contain. English demands number, an explicit subject in nearly every clause, and a gendered third-person pronoun. Japanese demands none of those, and instead requires you to encode your social relationship to the listener and the direction of favors, neither of which English tracks.

Translation between the two is lossy in both directions and lossy in different ways. Going in, the politeness and benefactive systems force choices the English never specified. Coming out, English forces number and subject choices the Japanese never specified.

Neither direction is a mistake by the machine. Both destroy information the target language cannot store. That is why the failures in this post are so hard to catch: there is no garbled output and no obvious error, just fluent text with something missing that you would have to know Japanese to notice.

Try It

The elephant sentence is a good one to start with, since the structure it breaks has no English equivalent at all:

→ Send Japanese through the mixer and watch the structure change

Sentences about objects come back clean. Sentences about people and social relationships go missing.

---

Related: We Round-Tripped One Paragraph Through 20 Languages has the full ranking Japanese topped. Finnish and Hungarian Don't Break Translators covers the languages at the other end of the table. And Why Machine Translation Humor Works explains why these particular failures are funny instead of just wrong.