Concept hub · ESL writing & AI detection
Translationese
Translationese is the statistical fingerprint of translated text: English where every sentence is correct, but the rhythm, word order, and phrase choices still follow another language. You do not need a translator to produce it — drafting in your head in one language and writing in another is enough. It is the single most common reason grammatically clean writing by non-native authors still reads as «off».
Where the term comes from
Translation scholars noticed decades ago that translated text is measurably different from text originally written in a language — not wrong, just different. Corpus studies (the line of work often traced to Gellerstam's 1986 study of Swedish novels translated from English) found consistent patterns regardless of translator skill. Those patterns got a name, and the name stuck: translationese.
The named components
Translationese is not one thing — it decomposes into patterns that each have a name:
- Calques — phrases translated word-for-word («make a photo», «according to me»). Every word is correct; the combination is borrowed.
- Explicitation — translated text over-explains: more connectors (moreover, therefore, in addition), more spelled-out logic than an English original would carry.
- Uniform rhythm — sentences of similar length and structure, because they were all produced by the same conversion process. English originals vary far more (what AI-detection research calls burstiness).
- Source-language interference — word order, article habits, and preposition choices that follow your first language. L1 interference is the mechanism; translationese is the surface result.
- Translation ghosts — sentences whose emphasis and information order still mirror the source language, like a ghost image showing through.
Check your own text
Paste a paragraph below. The checker scans the measurable patterns — passive density, sentence length, contractions, non-native phrase markers — and shows exactly what it found. It runs entirely in your browser; nothing is uploaded.
How to de-translationese a draft
Work through these moves in order — each one targets a named pattern:
- Split long sentences. Anything past ~25 words gets cut at the first «and», «which», or comma. This single move fixes rhythm faster than anything else.
- Turn the strongest passives active. «The team wrote the report», not «the report was written by the team». A little passive is fine; a lot reads converted.
- Allow contractions. Zero contractions across a page reads like output, not writing. A few natural «it's» and «don't» restore the human register.
- Hunt calques. Read each fixed phrase and ask: would an English writer reach for this combination? If unsure, the phrase probably came from your first language.
- Read one paragraph aloud. Rhythm problems are easier to hear than to see.
If you would rather not stop drafting to fix any of this, that is the workflow Diglot Weave was built for: write in the language you think in, translate inline, and let the editor flag the transfer patterns as you go.
Translationese and AI detectors
AI detectors score statistical uniformity: predictable word choices (low perplexity) and even sentence rhythms (low burstiness). Translationese has exactly those properties — which is a large part of why non-native writers get falsely flagged. A Stanford study (Liang et al., published in Patterns, 2023) found AI detectors falsely flagged 61% of TOEFL essays by non-native English speakers. If a detector has flagged your work, start with what to do when accused — and for work where it matters, the Authorship Certificate records your writing process as proof that exists before anyone asks.
Translationese — questions
What is translationese?
Translationese is the statistical and stylistic fingerprint of translated text: English that is grammatically correct but whose rhythm, word choices, and structures still follow the source language. The term comes from translation studies, where researchers found translated text differs measurably from text originally written in a language — more uniform sentences, more explicit connectors, word-for-word phrasings.
Why does my English have translationese if I did not use a translator?
Because self-translation counts. If you think a sentence in your first language and render it into English in your head — what linguists call mental translation — the output carries the same fingerprint as machine translation: source-language word order, calques, and rhythm. It is the default drafting mode for most non-native writers, especially under deadline pressure.
Is translationese the same as bad English?
No. Translationese is usually correct English — that is what makes it hard to fix with a grammar checker. Nothing is wrong at the sentence level; the text as a whole just does not read the way English originally written in English reads. The fixes are stylistic: sentence length, active subjects, contractions, idiomatic phrasing.
Do AI detectors mistake translationese for AI-generated text?
They can. Detectors flag statistical uniformity — even sentence rhythms, predictable word choices, low burstiness. Translationese has exactly those properties, which is one reason non-native writers are flagged far more often: a Stanford study (Liang et al., Patterns 2023) found AI detectors falsely flagged 61% of TOEFL essays by non-native English speakers.
How do I remove translationese from my writing?
Work top-down: split sentences that run past ~25 words, turn the strongest passives active, allow a few contractions, replace calqued phrases with the idiom an English writer would reach for, and read one paragraph aloud — rhythm problems are easier to hear than to see. The checker on this page shows which patterns your draft carries.
Diglot is a bilingual writing editor built for exactly this problem — start for free, no credit card required.
