What Is Translationese? Why Your English Sounds Translated

Translationese is the fingerprint that translation leaves on a text: the word choices, sentence shapes and rhythms that mark writing as translated rather than originally composed in that language. The term comes from translation studies, where researchers found that translated text is statistically distinguishable from native text even when every sentence is grammatically correct. Its main components have names: calques (phrases copied word for word from the source language), explicitation (spelling out what the original left implicit), source-language interference (grammar bleeding through) and an unusually uniform rhythm. And you do not need a translator to produce it. If you think in one language and write in another, converting as you go, your English carries the same fingerprint.

That last part is why the term matters far beyond translation departments. Most discussions of translationese are about published translations of novels and manuals. But the same features appear, for the same reasons, in the emails, essays and papers of anyone writing in a second language. The mechanism is identical; only the translator is different. This article covers both: what the term means in translation studies, what each component looks like in real sentences, why AI detectors sometimes misread translationese as machine writing, and how to remove it from a draft without sanding off your voice.

A term from translation studies, not an insult

Linguists have used “translationese” loosely for decades, but it became a research object in the 1980s, when corpus researchers began comparing large collections of translated text against text originally written in the same language. The Swedish linguist Martin Gellerstam, who popularized the term in a 1986 study, compared Swedish novels translated from English with novels written in Swedish, and found systematic differences in vocabulary and phrasing between the two. Not errors. Differences. The translations were fluent and correct, and still formed a measurably distinct variety of Swedish.

That finding has been replicated across many language pairs since, and it is the key thing to understand about translationese: it is not bad grammar. A text can pass every grammar check and still read as translated, because the fingerprint lives in distributions, not in individual mistakes. Which words you reach for, how often you make things explicit, how long your sentences run and how much that length varies. Some researchers go as far as calling translated language a “third code”: not quite the source language, not quite the target language, but a predictable blend with its own statistical habits.

For a second-language writer this framing is genuinely useful. It explains a frustrating experience: your text is correct, your teacher or editor cannot point to a rule you broke, and yet everyone agrees it “sounds translated”. Nothing is wrong. Something is different, and the difference has structure.

The named components, with examples

Calques

A calque is a phrase translated element by element from the source language, producing English that is understandable but not what English actually says. Classic examples:

Calques are the most visible layer of translationese and the easiest to fix once you spot them, because each one has a concrete replacement. The hard part is spotting them: to you, the calque feels natural, because it is your first language speaking through English. Close cousins are false friends, single words that look alike across languages but mean different things. Russian speakers, me included, have to unlearn “actual” — in Russian актуальный means relevant or current, so “this problem is very actual” keeps sneaking into drafts where English wants “this problem is very real” or “urgent”.

Explicitation

Explicitation is the tendency of translated text to spell out what the original left implicit. Translators add connectives, repeat referents instead of using pronouns, and unpack compact phrases into longer explanations. Translation researchers noticed in the 1980s that this happens across language pairs and directions, which suggests it is a property of the translation process itself, not of any particular language.

Second-language writers do the same thing for a related reason: fear of being misunderstood. If you are not sure your English carries the meaning, you add scaffolding. “Moreover”, “in addition”, “it should be noted that”, “as mentioned above” begin to open every second sentence. Each connective is fine alone. Stacked, they produce prose that explains its own structure constantly, the way a nervous tour guide narrates every turn. Native writing trusts the reader more and connects less.

Source-language interference

Interference is your first language’s grammar showing through English word order and structure. It is the deepest layer, because it operates below word choice:

• Articles. Slavic languages have no “a” and “the”, so Russian and Ukrainian writers drop articles or insert them by guesswork. Article errors are the most reliable single marker of a Slavic-language author I know, and I still make them when tired.
• Prepositions mapped from the source. “It depends from the context” (Russian зависит от), “we discussed about the plan”, “explain me this”.
• Word order. Languages that put known information first and news last produce English sentences that feel back-loaded: “About the results of this experiment we will speak later.”
• Passive and impersonal constructions. Academic conventions in many languages prefer “it was shown that” and “the analysis was performed”, and writers carry the habit into English at rates native writing does not sustain. A passive voice detector makes this layer visible quickly.

Uniform rhythm

The least discussed component and, statistically, one of the strongest. Corpus studies find translated text tends toward simplification: a narrower range of vocabulary, safer and more frequent words, and sentence lengths that even out toward the middle. Second-language writing shows the same pattern, and for an honest reason: you write with the words you are sure of. Risky idioms get avoided, sentence structures you trust get reused, and the result is prose with a very steady pulse. Correct, clear and strangely flat. Native writing is bumpier: a long winding sentence, then a short one. Fragments, even.

When you are the translator

Everything above was discovered by studying professional translations. But there is a private version of the same process that linguists call mental translation: composing the thought in your first language and converting it into English in your head, phrase by phrase, as you type. Almost every second-language writer does some of this, and everyone does more of it under pressure, fatigue or deadline. The output is self-translated text, and it carries the translationese fingerprint just as surely as a translated novel does.

English is my second language, and I can watch this happen in my own drafts. When I am fresh, I compose in English directly. When I am tired, the Russian sentence shapes come back: back-loaded word order, an impersonal construction, a calque I would catch in anyone else’s writing. The draft is still correct. It just stops sounding like me in English and starts sounding like me translated from Russian.

This is the exact problem Diglot’s weave mode exists for: instead of translating silently in your head and hoping the seams do not show, you write in both languages openly in one draft, and the tool handles the conversion where you want it. The translation becomes visible and deliberate rather than invisible and accidental.

Why detectors read translationese as AI-like

Here is where an academic term collides with real consequences. AI text detectors mostly measure two things: how predictable each word is given the previous ones, and how much the predictability varies across the text (see perplexity in AI detection). Machine text tends to be smoothly predictable. So the detectors look for smooth predictability.

Now reread the components above. Safer, more frequent vocabulary. Reused sentence structures. Evened-out sentence lengths. Extra connectives gluing everything into regular patterns. Translationese lowers exactly the statistics that detectors treat as evidence of machine authorship. The overlap is not a coincidence and not a conspiracy; it is two different processes, translation and language modeling, both producing more regular text than native composition does.

The consequences are documented. A Stanford study (Liang et al., published in Patterns, 2023) found AI detectors falsely flagged 61% of TOEFL essays written by non-native English speakers. There is also a stranger piece of evidence from the other direction: researchers led by Guillaume Cabanac found thousands of published papers containing “tortured phrases”, mangled synonyms such as “counterfeit consciousness” for “artificial intelligence”, left behind by paraphrasing tools. Processed text carries detectable marks, whether the processor is a translation habit, a paraphraser or a language model, and detectors are blunt instruments for telling those marks apart. If a flag has already landed on your work, we have a step-by-step guide for what to do about a false positive. And if you write with the constant low-grade fear of the next flag, that feeling has a name too: flagxiety.

How to de-translationese a draft

The fix is not “write more like a native”. It is a set of concrete passes, each targeting one component.

1. Hunt calques phrase by phrase. When a phrase feels slightly too comfortable, search it in quotation marks. If the results are mostly other second-language writers or nothing at all, English probably says it another way. Replace the phrase, then note it: your personal calques repeat, so each one you catch is caught for good.
2. Break the rhythm on purpose. Read the draft and mark sentence lengths. If everything runs 15 to 25 words, split one sentence into two short ones and merge two others into a long one. This feels artificial the first few times. It reads as natural, because native prose really is this uneven.
3. Cut the scaffolding. Delete “it should be noted that” everywhere; nothing is lost. Then allow yourself roughly one connective like “moreover” or “in addition” per paragraph, and let the remaining sentences sit next to each other without an usher.
4. Check prepositions after verbs. Depends on, discuss (nothing), emphasize (nothing), explain to me. Prepositions are where interference hides most reliably, because your first language chose them for you.
5. Convert impersonal passives back to actors. “It was decided that” becomes “we decided”. Keep passive where the actor genuinely does not matter; remove it where it is only an imported academic habit.
6. Read the final draft aloud, in English only. Mental translation cannot survive reading aloud at speaking pace; the back-loaded sentences and the calques stumble on the tongue before the eye ever catches them.

If you want the diagnosis before the surgery, we built a free sounds-translated checker for exactly this: paste a draft and it flags the phrases and patterns that read as translated, component by component, so you know which of the passes above your text actually needs.

One last thing, because the framing matters. Translationese is not a shameful condition; it is a visible stage of writing between two languages, and every bilingual writer passes through it in both directions. The goal is not to erase where you come from. It is to make sure that when your text reaches a reader, or a detector, it sounds like you writing English, not like you being converted into it.