- What is translationese?
- Translationese is language that is grammatically correct but reads as though it came from a translation rather than an original writer. The linguist Martin Gellerstam named it in 1986, describing it as a fingerprint the translation process leaves behind — not a sign of a bad translator, which is why it appears even in skilled and native work.
- How do I know if my Japanese sounds unnatural if I can't read it?
- You can't detect it through automated or functional QA — those check correctness, not naturalness. The only reliable layer is native-speaker review. Change the question your reviewers answer from "is this correct?" to "does this feel like a person wrote it?" and use more than one native reader.
TL;DR
Japanese copy can be completely correct and still read like a machine wrote it — a documented effect called translationese that leaves a "fingerprint" even on skilled, native work. Teams that cannot read Japanese have no way to detect it, because it passes every correctness, glossary, and functional check. The fix is not translating better but changing the question you ask — from "Is this correct?" to "How does this feel?" — and putting the copy in front of more than one native reader. That native-review layer is the only one that catches feel, and it is cheap to add.
Key Takeaways
- Correct and natural are two different targets. Japanese can pass every check and still sound non-human.
- Translationese (Gellerstam, 1986) is a fingerprint of the translation process, not a mark of a bad translator, which is why it reaches skilled and even native work.
- The real risk is detection. A team that cannot read Japanese cannot see it, no matter how good the dashboard.
- The naturalness gap lives in pacing — sentence length, comma placement, line breaks — not in vocabulary.
- The only reliable catch is native review that asks "how does this feel?" using more than one reader.
The sentence that still stings
A few years ago I localized the interface of a fintech product into Japanese. I am a native speaker. I read every string twice, checked the terminology, smoothed the phrasing. When I handed it off I genuinely thought it was clean.
The review was a shared meeting. Supervisors from other companies were in the room, which is a detail I remember more clearly than I would like. One of the Japanese reviewers read through my strings, paused, and said they "read like machine translation."
I did not have a good answer. The words were correct. The grammar was fine. Nothing was wrong in the way I had been trained to look for wrong. And yet a native reader had picked up something in a few seconds, in front of people whose opinion of me mattered.
That sentence still stings a little. I have kept it, because it turned out to be the most useful thing anyone has said about my work.
My Japanese wasn't wrong. It just didn't sound like a person.
Here is the uncomfortable part. I could not immediately fix it, because I could not immediately see it. My own copy looked normal to me.
So I went and read. Not translations. Real Japanese products, the ones people use every day without thinking about the language at all. Onboarding flows, error messages, empty states, buttons. I read them the way you read your first language, by feel, and I started noticing the gap between how those sentences moved and how mine moved.
Mine were accurate. They carried the meaning across. But they carried the shape of the English underneath them too. The rhythm was borrowed. A Japanese reader feels that borrowed rhythm the way you would feel a native English sentence that had quietly kept the word order of another language. Every word is a real word. It just does not sound like a person sat down and wrote it.
That was the lesson I did not want. Correct and human are two different targets, and I had only been aiming at one.
It has a name: translationese (since 1986)
The thing I had produced has a name. The linguist Martin Gellerstam called it translationese in 1986, describing the statistical fingerprint that translated text carries compared to text originally written in the same language (reference).
The word I want to hold onto is fingerprint. Translationese is not a synonym for a bad translator. It is a trace the translation process itself leaves behind, a residue of the source language showing through in the target. Researchers study it precisely because it appears in fluent, professional, native-quality work, not only in clumsy work (corpus research). That is why it reached mine.
This reframed the whole problem for me. I had been treating unnatural copy as a skill failure, something a better translator or a more careful pass would remove. But if it is a fingerprint of the process, then care alone does not lift it. You can be accurate and thorough and still leave the print behind, because the print comes from translating rather than from translating badly.
Which raises the real question for anyone shipping Japanese. Not "did we hire someone good enough," but "does our process have a step that catches the print." Most localization pipelines check correctness. Very few check for the fingerprint, because the fingerprint is invisible to everyone who cannot read the language by feel.
The three things I actually changed
Reading real Japanese taught me concrete habits. None of them are about vocabulary. All of them are about how a sentence lands.
Sentence length. English rewards a certain momentum. Clauses stack, ideas connect, the sentence keeps going. When I carried that momentum into Japanese, the result felt heavy and translated. Natural Japanese product copy is usually far shorter than the English rhythm suggests. I started cutting one long sentence into two or three. The meaning stayed. The machine feeling dropped away.
Where the comma sits. In Japanese, a comma is not only grammar. It controls where the reader's attention rests, which word gets the small beat of emphasis before it. Early on I placed commas by habit, roughly where the English had them. Once I started placing the comma right before the word I actually wanted to land on, sentences that had felt flat suddenly had a point of focus. Same words, different weight.
Where the line breaks. This one surprised me most. In a Japanese interface, text wraps, and where it wraps is not neutral. A line that breaks in the wrong place can make a perfectly correct sentence read as clumsy, because it splits a phrase the eye wants to keep whole. A line that breaks in the right place reads calm. In UI, where the screen is narrow and every string is short, the break point is part of the writing, not an afterthought handled by the layout.
You do not need to read Japanese to understand any of these. They are all the same underlying idea. Correctness lives in the words. Naturalness lives in the pacing, and pacing is exactly what a translated sentence quietly gets wrong.
Here's the trap: if you can't read Japanese, how would you ever know?
Now put yourself on the other side of my story. You are the team shipping the product. You do not read Japanese.
My strings would have passed every check you could run. The glossary matched. The grammar was valid. A spellchecker found nothing. A functional QA confirmed the text fit the buttons and the app did not break. And a native Japanese speaker wrote it, which is the reassurance most teams treat as the finish line.
Every one of those signals was green. The copy still read like a machine to the people it was for.
This is the trap, and it is structural, not a matter of effort. The one defect that damages you most in Japan is the one defect your tools are blind to, and the one your source-language team cannot feel. You can be diligent and still never see it, because seeing it requires reading the language the way a native reader does.
The cost is not a broken feature. It is quieter than that. Japanese buyers tend to read cautiously and weigh how serious and trustworthy a company looks before they commit (context, context). Copy that is subtly off does not throw an error. It just makes you feel slightly foreign, slightly less careful, at the exact moment the reader is deciding whether to trust you. That is a directional risk, not a number I can promise you. But it is the kind of thing that shows up as conversion you never earned and never knew you lost.
Change the question — from "Is this correct?" to "How does this feel?"
The fix that finally worked for me was not learning to translate better. It was changing the question.
"Is this correct?" is a question about the words. Tools answer it. Glossaries answer it. It is necessary, and it is not enough. The fingerprint survives a correctness check because the fingerprint is not an error.
"How does this feel?" is a question about the reader. It can only be answered by someone reading in their first language, reacting the way your actual customers will. Native-speaker linguistic review is the layer that catches feel, and it is the one layer automated and functional checks cannot replace (source). Not because the tools are weak, but because feel is not the kind of thing a tool is looking for.
The shift in one line: keep checking correctness, and add a step that checks how it lands — given to a human who reads the language by instinct.
What I do differently, cheaply, tomorrow
Two changes, neither expensive.
The first is that I show the copy to more than one native reader before it ships, and I ask them the feel question rather than the correct question. One reader gives you a data point. A second reader tells you whether the first reaction was personal taste or a real signal. It is a short review, not a second full translation, and it catches the fingerprint that every earlier step let through.
The second is more personal. I still run my own Japanese past other eyes. I am a native speaker who has spent years on exactly this problem, and I could not reliably see the print in my own writing in that meeting. I do not assume I can now. Being close to your own words is precisely what hides them from you, and that is as true for me as it was then.
So the cheap, boring, effective habit is this. More than one native reader, asked how it feels, before it goes live. It is what I do for my own work, and it is what we do at Hiraki.
A quieter question for your dashboard
You probably track whether your Japanese is correct. Most teams can see that number.
Here is the one most dashboards leave out. When was the last time more than one native reader looked at your live Japanese and told you, honestly, how it feels to read? If the answer is "we're not sure," that is not a failure. It is just the check you have not run yet. When you want a starting point, a Japan readiness check is built to look at exactly that.
Frequently Asked Questions
How is translationese different from a translation mistake?
A mistake is something wrong you can point to, like a bad word choice or broken grammar. Translationese is a fingerprint the translation process leaves even when nothing is wrong. It is about how the text feels, not whether it is correct.
We used a native Japanese translator, so are we safe?
Not automatically. Translationese appears in native and professional work because it comes from the act of translating, not from a lack of skill. Even native writers struggle to see the fingerprint in their own output, which is why a separate native review step matters.
Can machine translation or AI fix this?
Those tools mainly help with correctness and speed, and they can introduce their own translationese. They do not answer the "how does this feel?" question, which needs a human reading in their first language.
Why does this matter for conversion in Japan specifically?
Japanese buyers tend to read cautiously and weigh how trustworthy and serious a company looks before committing. Copy that is subtly off quietly undercuts that trust at the deciding moment. This is a directional risk, not a guaranteed number.
What is the cheapest first step?
Show your live Japanese to more than one native reader and ask how it feels to read, not whether it is correct. It is a short review, not a re-translation.