Most of the world writes English as a second language, often drafting in one language and typing in another. That writing tends to be careful and regular, which is exactly what AI detectors were built to treat as suspicious. This is a free book about why that happens, what Turnitin and the rest actually catch, and how to make a translation read like a person wrote it. Free to read here, free to share, no email required.
Plenty of people write in English without thinking in it first. They draft in Hindi or Spanish or Mandarin, or they think in one language and type in another, or they run a paragraph through DeepL and tidy it up. That is ordinary, and for most of the world it is simply how writing in English works.
It also produces text that AI detectors are unusually bad at reading. Translated and second-language English tends to be careful, regular and predictable, and those are exactly the properties detectors were built to treat as suspicious. The result is that some of the most honest writers get the worst scores.
This book is about that gap. What machine translation actually does to a sentence, what a detector sees when it reads one, what Turnitin and the rest can and cannot catch, and how to make a translation read like a person wrote it without pretending you did not use a tool. It is not a guide to slipping anything past anyone, and the chapter on academic policy says so plainly.
A machine translation is not really written in the target language. It is the source language wearing target-language words. The translator swaps vocabulary faithfully, but it tends to keep the shape of the original: its sentence order, its rhythm, its habits. Linguists call the result translationese, and it shows up as a handful of repeatable patterns.
Six of them do most of the damage:
Once you can name these, you can hear them, and once you can hear them, you can fix them. Let's go through each.
This is the big one. Every language has a default way of ordering a sentence, and a translator under pressure to be accurate often preserves the source order instead of rebuilding the sentence the way the target language prefers.
German, for example, parks verbs at the end of clauses. Spanish and French place adjectives after nouns and lean on long, comma-stitched sentences. Japanese drops subjects that English needs spelled out. When the translator keeps those habits, the English is technically correct but feels foreign.
Before: After the long meeting that lasted all morning had finally finished, decided we to take a break.
After: The meeting ran all morning. When it finally wrapped up, we took a break.
The first version keeps a borrowed clause order and an inverted "decided we." The second rebuilds it the way an English speaker would actually think it. Same meaning, completely different feel.
This is also why a single AI translation often reads more stiffly than the same sentence written fresh in English. The model is solving "say this in English," not "what would an English writer have said here."
A translation model is trained to pick the most probable next word. That is great for not being wrong, and terrible for sounding like a person. Real writers reach for the specific word, the slightly surprising one, the verb that carries an attitude. Machines reach for the average one.
The effect is a kind of beige fluency. Nothing is incorrect, and nothing has any flavor.
Before: The results were very good and showed important things about the problem.
After: The results were striking. They exposed a flaw nobody had flagged.
"Very good," "important things," and "the problem" are all safe defaults. "Striking," "exposed," and "a flaw nobody had flagged" are choices. A human picks; a model averages. That gap is a big part of why AI translations sound robotic even when every word is defensible.
Humans get bored of their own words and switch them up. Machine output does not get bored, so the same connector, the same sentence opener, and the same go-to noun keep returning. You notice it as a flatness, a rhythm that never changes gear.
Before: It is important to note that the study is important. It is also important to consider the limitations, because limitations are important to the conclusion.
After: The study matters, but its limitations shape what we can take from it.
Three "importants" in two sentences is the tell. Native writing varies sentence length too, mixing a short punchy line with a longer one. Machine translations tend to march out sentences of similar length, which is part of why the whole thing feels mechanical. This evenness is also, not coincidentally, one of the traits that automated detectors key on, which we cover in the translationese detection guide.
Idioms are where literal translation falls apart fastest, because the meaning lives in the phrase, not the words. Translate the words and you get something that is correct on paper and baffling in practice.
A French speaker who says the equivalent of "I have the peach" means they feel great. A literal English render of that is nonsense. The same goes the other way: "it's raining cats and dogs" translated word for word into most languages produces a small surreal scene.
Before: We need to take the bull by the horns and put the church back in the middle of the village.
After: We need to face this head on and get our priorities straight.
The second clause there is a real French idiom rendered literally, and it means roughly "restore order." Translated word for word it just confuses an English reader. Good translation swaps the idiom for an equivalent one in the target language, or drops the figure of speech and states the meaning plainly.
False friends are words that look or sound nearly identical across two languages but mean different things. They are a classic source of translations that read as slightly wrong in a way readers cannot always pinpoint.
Spanish "actual" means current, not actual. French "éventuellement" means possibly, not eventually. German "Gift" means poison. "Embarazada" in Spanish means pregnant, not embarrassed. A model that trusts the surface resemblance produces a sentence that is grammatically fine and quietly incorrect.
Before: The author actually argues that the policy is sensible and the result was eventually positive.
After: The author currently argues the policy is sensible, and the result was ultimately positive.
If "actually" was meant to render Spanish "actualmente," it should be "currently," and "eventually" for French "éventuellement" should be "possibly" or, depending on context, "ultimately." These slips rarely break a sentence outright. They just make it feel subtly off, which is often more damaging because the reader trusts it less without knowing why.
Register is the formality setting of language: a text message and a cover letter are the same English in very different registers. Translation models are weak at holding a consistent register, partly because the source language may not split formal and casual the same way English does, and partly because the model has no strong sense of who is speaking to whom.
The result is tone that wobbles. A casual blog post picks up stiff, legalistic phrasing in one paragraph, then drops into something oddly chatty in the next.
Before: Hey folks, herewith we shall elucidate the matter at hand, so don't sweat it, the aforementioned points are pretty chill.
After: Hey folks, here's the breakdown. Don't worry, these points are simple.
The "before" mixes "hey folks" and "don't sweat it" with "herewith," "elucidate," and "aforementioned." No human writes both in one breath. A clean translation picks one register and holds it the whole way through.
The good news: once you know the six patterns, fixing them is mostly mechanical. You are not rewriting from nothing, you are correcting known faults. Here is the by-hand version.
For a full worked method with examples at each step, see how to make ChatGPT and DeepL translations sound natural.
Doing all six by hand is the highest-quality route, and it is the right one when the stakes are high. But when you have a long passage and limited time, a humanizer does the same job faster by rewriting for natural flow while preserving your meaning.
The honest framing matters here, so we will be plain about it. A humanizer is a writing-quality tool. It is for turning a robotic translation into clear, natural prose that says what you meant. It is not a way to disguise anything or get around a check, and we do not pitch it that way. Used to make your own translated writing read well, it is genuinely useful, and that is the whole point.
The chapter on fixing a stiff translation covers that workflow end to end.
Start with the thing detectors are reacting to. It has a name, and once you can hear it you can hear it everywhere.
At its simplest, detection answers one question: did a machine do the translating? There are two broad ways to answer it, and they are not the same skill.
Most real situations use both. Turnitin, for example, runs an automated similarity and AI check, then a human grader looks at anything the system flags. We cover how that specific pipeline handles translated work in Can Turnitin detect translated text? The important point for now is that detection is always an estimate of likelihood, never a confession. No tool and no reader can prove that a machine wrote something from the text alone.
When a machine translates a sentence, it tends to keep the shape of the original language and swap in the nearest English words. The grammar comes out correct, but the rhythm and word choices stay loyal to the source rather than to how a native speaker would have built the thought from nothing. Linguists call the result translationese, and it is the single most useful concept in this whole topic.
Translationese is not the same as bad writing. It can be perfectly clear and grammatically clean. What gives it away is a slight stiffness, a sense that the sentences were assembled rather than spoken. The reason machine translations feel this way is worth understanding on its own, and we walk through the mechanics in why AI translations sound robotic.
These are the tells that both human readers and automated tools lean on. No single one proves anything, but together they form a pattern.
The next chapter gives an example-driven breakdown of these signals and how to weigh them. If your specific question is whether one particular passage came from a tool, is this translation AI-generated? walks through a quick test.
Automated detectors do not read for meaning the way you do. They measure the statistical fingerprint of the text. The core idea is predictability: human writing varies, with some sentences running long and word choices that wander, while machine output is smoother and more uniform. That uniformity is what a detector keys on.
Here is the overlap that makes this whole field tricky. The traits that mark translationese, even phrasing, predictable word choice, low variety, repetitive rhythm, are almost exactly the traits an AI detector measures. So machine-translated text and machine-written text look alike to a detector, which is why one tool ends up flagging both. It is also why the error rate climbs on translated content: a careful human writing in their second language can share some of those surface traits without any tool involved, and the detector cannot tell the difference from the numbers alone.
If you want the dedicated tool for this exact job, our AI translation detector is built around the translated-text case specifically, and it is honest about what it can and cannot catch.
This is where most of the marketing around detection tools oversells, so it is worth being plain.
So the accurate framing is not "a detector always catches machine translation." It is "a detector often flags machine translation, with meaningful error in both directions, and more error on translated text than on native writing." That honesty is the point. A tool that pretends to certainty it does not have will eventually accuse the wrong person.
Detection is not all-or-nothing, and which tool produced the translation shifts how strong the signal is.
The near-paradox is worth holding on to: the more natural a raw machine translation sounds, the more it can read like polished AI text to a detector. So switching translators does not reliably change the outcome.
Automated tools are useful, but two human methods remain some of the most reliable ways to catch machine translation. Neither needs software.
Take the English text and translate it back into the suspected source language with a tool, then compare. If the original was machine-translated, the round trip often snaps back surprisingly close to a known source, or it exposes the literal idioms and odd word order that a human translator would have smoothed over. Back-translation will not give you a clean yes or no, but it is one of the best ways to surface translationese that a quick read might miss.
Slow reading still beats most tools in a language you know well. Look for the idioms that land flat, the register that drifts, the small connecting words that pile up, and the sentences that all breathe at the same length. A human reader who knows both languages can often feel that a passage was built rather than written, even when every individual sentence is correct. This is the skill behind a teacher's instinct, and it is why human review stays in the loop even where automated detection runs first.
If you have text in front of you, there are really only two useful moves, and they map cleanly to the two sides of this work.
Check it. Run the passage through an AI detector to get a probability score and sentence-level highlights, so you can see which lines read as machine-smoothed. Used as a guide for your own revision, on your own work, before anyone else sees it, that is genuinely helpful. Just keep the limits above in mind, especially on translated text. The detector is a second opinion, not a final word.
Naturalize it. If the goal is to take a stiff, word-for-word translation and make it read the way a person would actually write, that is a rewriting job. Our AI humanizer is built to fix translationese while keeping the original meaning, and the dedicated walkthrough lives at humanize translated text. For a hands-on method you can apply yourself, making AI translations sound natural lays out the steps: match the register, replace literal idioms, and vary the rhythm.
One thing this is not about: getting past a checker. If your situation is academic, the question is rarely technical and usually about policy, so it is worth knowing what actually counts. We cover that neutrally in is translating your essay with AI cheating? The honest path is always to check and improve your own work, never to disguise it.
Translationese is the general shape. These are the specific tells, the ones you can point at in a sentence.
Before the signs, set your expectations. You are reading clues, not running a test that returns a yes or no. A skilled human translator can produce writing that looks a little stiff, especially under deadline, and a strong AI translation from a tool like DeepL or ChatGPT can read smoothly enough to pass for human on a good day. The signs below stack the odds. They do not prove authorship.
That matters most when something is at stake, such as grading a student, reviewing a freelancer, or assessing a submission. Treat any one sign as a reason to look closer, not a verdict. The more signs you find, and the more they cluster, the more confident you can be. With that framing in place, here are the seven.
The single biggest tell is translationese. The grammar is correct, but the English keeps the bones of the original language instead of flowing the way a native speaker would write from a blank page. The result reads "off" in a way that is hard to name until you see it side by side.
Take a French source. A human might write, "She finally agreed." A machine often produces something closer to, "She has finished by accepting," because that mirrors the French construction word for word. Nothing there is ungrammatical. It is just not how an English speaker would say it. When a whole document carries that faint accent of another language, you are almost certainly looking at machine output that nobody rewrote afterward. A human translator's job is to break that mirror. A tool, left alone, does not.
Idioms are where machine translation falls hardest, because an idiom means something other than its words. A good human translator swaps the idiom for an equivalent one or rephrases the idea. A machine tends to render it literally, and the result lands flat or sounds faintly absurd.
The German "Ich verstehe nur Bahnhof" means "I don't understand a thing." Translated literally you get "I only understand train station," which no English speaker would ever write. Most cases are subtler than that, but the pattern holds: a phrase that is almost an English saying but not quite, or a metaphor that does not sit right, often points back to an idiom in the source language that the tool carried across word for word. Humans localize idioms. Machines transcribe them.
This sign is quieter and more statistical, which is exactly why it is useful. Machine translation leans on small function words, the, of, in, to, that, far more than natural English does. The source language often requires those connective words, and the tool dutifully keeps every one rather than dropping the ones English would leave out.
Compare two versions of the same idea. A machine might write, "The analysis of the results of the study of the group." A human would tighten that to, "The group's study results." Same meaning, far fewer little words. When you notice a paragraph clotted with of and the, where an editor's instinct is to cut half of them, that density is a real signal. It is also one of the patterns automated detectors measure, which is why this sign and the tools tend to agree.
Human writers reach for variety almost without thinking. They swap synonyms, vary their phrasing, and avoid repeating the same word three times in a paragraph unless they mean to. Machine translation, by contrast, tends to pick the single most statistically likely word for a concept and then reuse it every time that concept appears.
So you get a passage where "important" shows up in four sentences running, or where the same connective ("however," "moreover") opens paragraph after paragraph. A human editor would have varied it on instinct. The flatness is not a sign of a bad writer. It is a sign of no writer, because the tool has no preference for variety the way a person does. Read a long stretch and count the repeats. Narrow, repetitive word choice across a whole document is a strong tell.
Beyond word choice, listen to the rhythm. Human writing breathes unevenly. A long, winding sentence is followed by a short one. A fragment lands for effect. Sentence length wanders because thought wanders. Machine translation flattens that rhythm into a steady, even pulse, with most sentences landing at a similar length and built on a similar frame.
Read a suspected passage out loud, or just scan the sentence lengths down the left margin. If almost every sentence is medium length and structured the same way, subject, verb, object, clause, with little variation, you are hearing the machine's even hum rather than a human's natural cadence. This evenness is one of the core things AI detectors are tuned to catch, under the name low burstiness, and it is something you can train your own ear to notice without any tool at all.
A human translator holds a consistent voice across a document, because one person is making every choice. Machine translation works in chunks and has no memory of the voice it set a paragraph ago, so the register can drift without reason. A formal report suddenly turns casual for a sentence. A friendly email goes stiff and bureaucratic in the middle. A technical manual slips into oddly conversational phrasing and then back again.
These shifts are jarring precisely because nothing in the content explains them. A human writer changing tone usually does it on purpose and signals it. When the tone lurches with no purpose, especially across paragraph boundaries, it often means the passage was translated piece by piece by a tool that never held the whole thing in view. Watch for the seams where the voice changes for no reason.
The last sign shows up in specialized writing. Every field has its own settled vocabulary, the exact term a practitioner uses, and a human expert reaches for it automatically. Machine translation often misses that established term and produces a literal, dictionary-correct phrase instead, one that means the right thing but no insider would ever say.
In a legal document, a tool might render a concept as "the right of going back" where the field's actual term is "right of recourse." In medicine, software, or finance, you see the same thing: the meaning is intact, but the term is the descriptive long-form version rather than the crisp term of art. To a layperson it reads fine. To anyone who works in the field, it instantly signals that whoever produced this was not a domain translator, and very possibly not a human one. When the ideas are right but the terminology feels translated rather than known, that is your seventh sign.
Spotting signs is the first half. Confirming them is the second, and there are two practical checks worth running before you draw a conclusion.
The simplest manual check is back-translation. Take the suspect text, run it through a machine translator into the language you think it came from, and read the result. If it snaps neatly back into fluent, idiomatic text in that language, the structure was probably built for that language in the first place, which points to machine translation from that source. Genuinely human English, written from scratch, tends to come apart awkwardly when you push it back the other way, because it was never shaped around the source grammar to begin with. It is not foolproof, but it is free, fast, and often clarifying.
The other check is an AI detector, which measures the same statistical evenness, the low burstiness and predictable phrasing, that signs four and five describe. Running the text through a detector gives you a probability score and, in good tools, sentence-level highlights showing which lines read as machine-generated. That turns a vague hunch into something you can point at.
Here is the honest limit, and it is the reason we say verify rather than prove. AI detectors are noticeably less reliable on translated text than on text written directly by a chatbot. Published testing has found accuracy can drop by roughly twenty percent on machine-translated content, because translation muddies the very signals the detector relies on. Detection is probabilistic in the first place, so a score is a confidence estimate, not a confession, and false positives are real. They fall hardest on writers working in a second language, whose careful, even English can look machine-made when it is not. So use a detector as one input alongside the seven signs and the back-translation read. Never use it alone, and never treat a number as the final word. For the full linguistic background on why translated text confuses these tools, our opening chapters on machine translation detection and translationese goes deeper.
No single sign is conclusive. That is the whole point. A repeated word might be a stylistic choice. One flat idiom might be a tired human. One tonal wobble might be a quick edit. The method is to read for all seven and see how many show up together. One sign is a shrug. Two or three that cluster, say translationese plus narrow vocabulary plus an unexplained tone shift, start to tell a clear story. Five or six, confirmed by a back-translation that snaps cleanly into the source language and a detector that flags the same passages, is about as confident as you can honestly get without a confession.
Knowing the signs is not the same as making a call on a particular piece of text. This chapter is for the moment you actually have to decide.
Before any tool, trust your ears. Read a paragraph out loud. Human writing, even translated human writing, has a rhythm. It speeds up, slows down, stumbles in a way that sounds like a person thinking. AI translation tends to move at one even pace. Every sentence is roughly the same length. Every clause lands in the same place. Nothing surprises you.
That flatness is the single most reliable first signal. It is not proof, and you should never stop here, but if a passage reads like a calm, well-mannered robot that never gets excited or distracted, you have a reason to look closer. Make a note of which paragraphs gave you that feeling. You will test those first.
Once you have a candidate passage, look for specific patterns rather than a vague "it feels weird." Here are the tells that actually mean something. None of them is a smoking gun on its own. Two or three together start to matter.
Literal idioms. Idioms are where machine translation gives itself away most often. A phrase that is natural in the source language gets carried across word for word, so it lands flat or slightly wrong in the target language. "It costs an arm" instead of "it costs an arm and a leg." "Make attention" instead of "pay attention." A human translator would smooth these into the local equivalent. AI frequently does not, because it followed the source structure too closely.
Low word variety. Count how many times the same connective words show up: however, moreover, furthermore, in addition, therefore. AI translation, like AI writing in general, reaches for the same safe vocabulary again and again. A native writer reaching for variety would mix it up. Uniform word choice across a whole document is a tell.
Tone that shifts for no reason. Watch for a sentence that is suddenly more formal or more casual than the ones around it. Machine translation handles each sentence in a fairly isolated way, so the register can wobble. A paragraph that drifts from chatty to bureaucratic and back, without the author meaning anything by it, is a sign the text was assembled rather than written.
Source-language grammar showing through. This is the deep one. Translated text often keeps the sentence shape of the original language even when the words are correct. Adjective order, where the verb sits, how long the run-up to the main point is, all of it can carry the structure of the source. Linguists call the result translationese. The chapter on the seven signs breaks that full list down.
Over-explained or oddly generic phrasing. AI tends to translate cautiously, adding little clarifications a human would trust the reader to get, or flattening a colorful original into something safe and plain. If a passage feels like it lost all its personality somewhere along the way, that is worth noting.
This is the one test in this chapter that is worth doing every time, because it turns a hunch into something you can actually see. It is simple and it costs nothing.
Take a suspicious passage and translate it back into the source language using a free tool like Google Translate or DeepL. Then read the result against the original source, if you have it, or just read it on its own for naturalness.
Here is what you are looking for. When something was machine-translated, running it back through a translator tends to produce text that snaps neatly back toward the original phrasing, because the machine output stayed so close to the source structure in the first place. Human translation, by contrast, takes liberties. A person rewrites for the target audience, so when you back-translate it, you get noticeable drift. Words change. The structure loosens. It does not reassemble cleanly.
So the rough rule is this. Little drift on back-translation points toward machine translation. A lot of drift points toward a human. A person made choices a tool would not, and those choices do not reverse neatly.
A few honest cautions, because this test is useful but not magic:
Used together with the manual tells, the back-translation test is the strongest thing you can do by hand. If the read-aloud felt flat, the idioms landed literally, and the passage snapped back cleanly to the source, you have three independent signals pointing the same way. That is a reasonable basis to look harder, not yet a verdict.
After the manual checks, an AI detector is a useful second opinion. It looks at the same kind of evenness your ear noticed, but it measures it across the whole document and gives you a probability score with sentence-level highlights. That can confirm a hunch or flag a passage you skimmed past.
Now the honest part, because this is where most guides oversell. Detectors are less reliable on translated text than on text written directly in the target language. Published testing has found accuracy can drop by roughly twenty percent on translated content compared with native AI writing. The reason is the overlap we keep coming back to: machine translation and AI writing share the same surface traits, even phrasing and predictable word choice, so a detector tuned to spot AI writing gets confused by translation in both directions. It can miss machine translation, and it can wrongly flag careful human translation that happens to read evenly.
Two things follow from that, and they are worth keeping in front of you:
So the right way to use a detector here is as one voice in the room, not the judge. It confirms or complicates what your manual checks already suggested. If the read-aloud, the idioms, the back-translation, and the detector all point the same way, your confidence is high. If they disagree, the disagreement itself is information, and the answer is to look more, not to trust the tool over your own eyes. The chapters on Turnitin, Google Translate and DeepL give a fuller account of what detectors can and cannot do on translated material.
Deciding "this is probably AI" is not the end. What you do with that depends on who you are and why you were checking.
If you are a teacher or grader. Do not treat a detector score as evidence on its own, and do not accuse based on a single tool. Use what you found as a reason to start a conversation. Ask the student to walk you through their drafts, their sources, and their process. Genuine work leaves a trail: version history, notes, a translated source they can cite. The presence or absence of that trail tells you far more than any score. If your institution uses Turnitin, the chapter on what Turnitin actually checks explains what those checks do and how reliable they are.
If you are an editor or client. Go back to the writer with specifics rather than a verdict. Point at the flat paragraphs, the literal idiom, the tone wobble. Ask them to rewrite those sections in their own voice. Whether a tool was involved or not, the fix is the same: the writing needs the human pass it did not get. You are improving the work, not running a trial.
If you are just a reader trying to decide whether to trust something. Weigh the stakes. For a casual email, a flat machine translation usually does not matter. For a contract, a medical instruction, or anything where a mistranslated nuance could cause harm, the right move is to get a human translator to review the original. AI translation can be confidently wrong in ways that read perfectly smooth, and smoothness is exactly what you cannot rely on.
In every case, the goal is not to win a gotcha. It is to make a sound decision about whether the translation can be trusted as it stands, and to know what would make it trustworthy if it cannot.
If you want the whole thing as a checklist you can run on one passage:
If you want to go deeper on the linguistics behind all of this, the opening chapters guide on machine translation detection and translationese collects the full set of signals and explains why translated text and AI text overlap in the first place.
Now the question most people arrive with. It has a real answer, and it is more limited than either the worried student or the confident vendor tends to assume.
Turnitin works in two layers, and translated text can trip either one.
Neither layer reads your mind. Both return a probability, not a verdict. That distinction matters, and we come back to it below.
For years the simplest way to dodge a plagiarism checker was to copy a source in one language and translate it into another. Turnitin closed most of that gap with cross-language matching. When you submit a paper, the system can translate it back into the languages in its index and look for the original. If your "own" paragraph is really a translated copy of a published article, the match still shows up.
This is why translating a source does not make it original. The words changed, but the structure, order of ideas, and specific claims usually did not, and that is what the comparison catches.
The newer piece is AI detection. Turnitin's model was first trained on output from GPT-3 and GPT-3.5, then expanded to recognize patterns from GPT-4, Gemini, Llama, and similar models. It looks at how predictable each sentence is. Human writing tends to vary: some sentences run long, some are short, word choices wander. Machine output is smoother and more uniform, and that uniformity is the tell.
Translation tools create the same kind of smoothness. When DeepL or ChatGPT rewrites a sentence in English, it tends to pick the most statistically likely phrasing, which is exactly the pattern the detector is tuned to notice. Turnitin also added a check for AI paraphrasing, so text that was AI-written and then run through a spinner to "rough it up" can be flagged as AI-generated text that was later AI-paraphrased.
Here is where honesty matters, because the marketing around detection tools tends to oversell.
Turnitin reports that it flags a large share of fully AI-written English text and claims a low false-positive rate on clean human writing. Independent reviews tell a more careful story. Independent testing of detection tools put real-world accuracy closer to the mid-70s percent range, not the high 90s, and accuracy drops further on translated content, by roughly a fifth (around 20 percent) in published testing.
Two things follow from that:
So the accurate statement is not "Turnitin always catches translated text." It is "Turnitin often flags translated and AI-translated text, with meaningful error in both directions." Anyone who tells you a detector is a perfect lie-detector is selling something.
Detection is not all-or-nothing. The translator you used changes the odds.
The pattern is almost a paradox: the more natural and "human" a raw machine translation sounds, the more it can look like polished AI text to a detector. That is why simply switching translators is not a reliable way to pass.
It helps to understand the linguistic reason, because it also tells you how a human reviewer spots translation by eye.
Machine translation produces what linguists call translationese. The grammar is correct, but the writing keeps the shape of the source language instead of flowing the way a native speaker would write from scratch. Common signs include overused function words (the, of, in), low vocabulary variety, repetitive sentence rhythm, and idioms translated too literally so they land flat.
Those same traits, even phrasing, predictable word choice, low variety, are what an AI detector measures. So translated text and AI-written text overlap, which is why one tool ends up flagging both, and also why the error rate is higher: genuinely careful human writing in a second language can share some of those traits without any tool involved. The chapter on translationese gives the full breakdown of those signals and how to spot them by eye.
Plenty of people translate for honest reasons. You drafted in your first language and moved it to English. You quoted a foreign-language source and translated the quote. You used a tool to clean up grammar. None of that is misconduct on its own, but it can still trip a detector, so a little care protects you.
The most useful move is to see what a detector sees before anyone else does. Running your text through an AI detector first shows you which passages read as machine-generated, so you can revise the weak spots rather than find out after the fact.
If the goal is to make a stiff, machine-translated passage read naturally again while keeping your meaning, that is a rewriting job, and the opening chapter walks through how to fix it by hand.
Google Translate deserves its own chapter, because it is the tool most people actually use and it behaves differently from a chat model.
Google Translate gets caught through three doors, not one.
None of these reads your mind. Each returns a probability or a match percentage, not a confession. We come back to what that means below.
For a long time the oldest trick in the book was to copy a source in one language and translate it into another to dodge the plagiarism checker. Turnitin closed most of that gap with cross-language matching. When you submit, the system can translate your text back into the languages in its index and look for the source it came from.
Google Translate is especially exposed here because it is faithful to a fault. It tends to follow the source sentence almost word for word, keeping the same order of ideas and the same specific claims. That faithfulness is exactly what cross-language matching keys on. So a paragraph that started as a published Spanish or German article, run through Google Translate into English, often still lines up with the original well enough to show as a match. The translation did not make it original. It just changed the surface.
The newer piece is AI detection. Turnitin's model learned the patterns of machine-written English from GPT-3 onward, and it looks at how predictable each sentence is. Human writing varies in length, rhythm, and word choice. Polished machine output is smoother and more uniform, and that evenness is the tell.
Here is the wrinkle with Google Translate. Its output is not always smooth and confident the way a chat model's is. It can be choppy, literal, and a little broken. That roughness means a pure AI writing check sometimes scores Google Translate lower than it scores ChatGPT, because the text does not read like fluent, statistically average prose. People take this to mean Google Translate is safe. It is not. The weaker AI score is more than made up for by the similarity match and by how obvious the clunky phrasing is to a person reading it.
This is the part that surprises students. You might assume the simplest, most popular translator is the hardest to catch. The opposite is closer to the truth.
So there is almost a trade-off baked into each tool. Google Translate loses on the two checks that catch the most students in practice, the similarity match and the human read, even if it sometimes wins on the automated AI score. The smoother tools win on readability and lose on the AI fingerprint. Switching from one translator to another does not make text safe. It just moves which door it gets caught at. The chapter on translationese covers the linguistic reason behind all of this.
This matters, because the marketing around detection oversells, and the panic around it overcorrects.
Turnitin reports that it flags a large share of fully AI-written English and claims a low false-positive rate on clean human writing. Independent testing is more cautious. A 2025 review of detection tools put real-world accuracy closer to the mid-70s percent range rather than the high 90s, and accuracy drops further on translated content, by roughly a fifth in published testing. Translated text is genuinely one of the harder cases for the automated check.
Two things follow from that:
So the honest statement is not "Turnitin always catches Google Translate." It is "Turnitin usually catches Google Translate, most often through the similarity match and the obvious phrasing, with real error in both directions on the AI score." Anyone who promises a detector is a perfect lie-detector is selling you something.
Software is only half the story. A grader who has read a few hundred papers develops an ear, and Google Translate sets off that ear faster than the smoother tools.
The reason is translationese: text that is grammatically fine but keeps the shape of the source language instead of flowing the way a native speaker writes from scratch. With Google Translate you tend to see literal idioms that land flat, oddly formal or stilted word choices, repetitive sentence rhythm, and the occasional phrase that is technically correct but no human would ever say. Add in a tone that does not match the rest of the paper, and a grader notices well before any score appears.
That human read is why Google Translate is risky even on assignments that are never run through software. The tell is in the text itself, not only in the report.
Plenty of people reach for Google Translate honestly. You drafted in your first language and moved it to English. You quoted a foreign-language source and needed the gist. You wanted to check your own grammar. None of that is misconduct on its own, but raw Google Translate output can still trip a detector and a grader, so a little care protects you.
The pattern that gets people in trouble is paste-and-submit. The pattern that keeps you safe is translate, then genuinely rewrite, then cite what you borrowed.
The most useful move is to see what a detector sees before anyone else does. Running your text through an AI detector first shows you which passages read as machine-translated, so you can revise the weak spots instead of finding out after the grade.
If your goal is to take a stiff, word-for-word Google Translate passage and make it read like a person wrote it while keeping your meaning, that is a rewriting job. Our companion guide on whether Turnitin can detect translated text covers the wider set of translators if you used more than one.
DeepL has a reputation for sounding more natural than its competitors. That reputation is partly earned, and it changes what a detector sees.
DeepL is a machine translation engine, not a chatbot, but for detection purposes that distinction matters less than you would think.
So the honest framing is not "DeepL always gets caught" or "DeepL is safe." It is "DeepL output often reads as machine-made to a detector, with real uncertainty in both directions."
AI detectors work by measuring how predictable a piece of writing is. They look at two related signals: perplexity, which is roughly how surprising each next word is, and burstiness, which is how much sentence length and rhythm vary across a passage. Human writing tends to be bursty and a little unpredictable. We write a long sentence, then a short one. We reach for an odd word now and then. Machine output is smoother and more uniform, and that uniformity is the fingerprint detectors are tuned to find.
DeepL is built to produce fluent, idiomatic English, which means it tends to choose the most natural and statistically likely phrasing at each step. That is great for readability. It is also, almost word for word, the description of what a detector flags. The model picks safe, common, well-formed phrasing, and a detector reads that consistency as a machine signal.
This is the core paradox. A clunky, literal translation full of awkward phrasing might actually score lower on an AI check, because the awkwardness reads as messy and human, even though a human grader would spot it as machine work in a second. A polished DeepL translation can score higher on the AI check precisely because it is clean. Quality and "detectability" pull in the same direction here, which is the opposite of what most people expect.
Turnitin runs two separate checks, and DeepL output can touch both. Its AI writing check looks for the machine signature described above, and DeepL's smooth phrasing can register there. Its similarity check translates submissions back into other languages and compares them against its database, so a DeepL translation of an existing published source can still match the original it came from. We cover the academic side in depth in can Turnitin detect translated text, but the short version is that switching to a better translator does not make a copied source original.
Here is where the honesty has to come first, because the marketing around detection tools tends to oversell.
Independent testing has found that AI detectors lose accuracy on machine-translated text, with published reviews reporting roughly a 20 percent drop compared with native English content. That happens for a specific, understandable reason, and it cuts both ways.
Machine translation produces what linguists call translationese. The grammar is correct, but the writing keeps the shape of the source language instead of flowing the way a native speaker would write from scratch. Common signs include overused function words like the, of, and in, low vocabulary variety, repetitive sentence rhythm, and idioms translated too literally so they land flat. Those traits, even phrasing, predictable word choice, low variety, are the exact same things an AI detector measures.
So translated text and AI-written text overlap heavily, and that overlap is why the error rate climbs:
The takeaway is not to distrust detectors entirely. It is to treat a score on translated text as a softer signal than a score on native writing. The chapter on translationese breaks down these linguistic tells and how to catch them by eye.
Not all machine translation looks the same to a detector, and the differences are worth knowing.
The pattern across all three is the paradox again: the more natural a raw machine translation sounds, the more it can look like polished AI text to a detector. Switching from Google Translate to DeepL improves the reading experience, but it does not reliably change whether a detector flags the result.
If you have a DeepL translation and you care how it reads, there are two useful moves. Neither is about gaming anyone. Both are about understanding and improving your own work.
Before you hand anything in or publish it, see what a detector sees. Running your DeepL translation through an AI detector first shows you which passages read as machine-generated, so you can revise the weak spots rather than find out after the fact.
The deeper fix is not to hide the translation but to make it read the way you would actually write. A raw DeepL passage often carries translationese: stiff rhythm, phrasing that mirrors the source language, word choices that are correct but not how you talk. Smoothing that out makes the writing genuinely better, and it happens to reduce the machine evenness a detector measures, because natural human variation is exactly what was missing.
To be clear about what this is and is not: naturalizing a translation means making it read like a person wrote it, in your own voice, with your meaning preserved. It is not a trick to defeat a detector, and any tool that promises to make writing "undetectable" is selling you a result it cannot honestly guarantee.
Plenty of people use DeepL for honest reasons. You drafted in your first language and moved it to English. You quoted a foreign-language source. You wanted cleaner grammar than you could manage in a second language. None of that is misconduct on its own, but it can still trip a detector, so a little care protects you.
Everything so far has been diagnosis. The rest of the book is the fix, starting with the method that produces the best result and takes the most effort.
A good edit changes how the sentence lands, not what it says. You are not rewriting the content, adding ideas, or softening claims. You are taking a sentence that is technically correct and making it read the way a fluent speaker would have written it from scratch.
Keep the original next to your draft while you work. Every time you change a line, check it against the source meaning. If a sentence sounds better but now says something slightly different, you have gone too far. The whole point of translation is fidelity, so the natural version has to carry the same information as the literal one.
The six steps below move from the broadest fix to the finest. Do them in order. Register first, because it sets the tone for everything else, and the read-aloud check last, because it catches whatever the earlier steps missed.
Register is the formality level: casual, neutral, or formal. Machine translation tends to flatten everything toward a stiff middle. A friendly product email comes out sounding like a contract, and a legal clause can come out weirdly chatty. Before you touch individual words, decide what register the original actually had, then push the whole passage toward it.
Read the source and name the tone in one word. Warm. Technical. Blunt. Then read your translation and ask whether it matches. Usually it does not, and the gap is the first thing a reader feels even if they cannot name it.
Before: "We are pleased to inform you that your request has been received and shall be processed in due course."
After: "Thanks for reaching out. We have your request and we will get to it soon."
Both say the same thing. The second one matches a casual support email. If the source had been a formal notice, you would move the other direction, but you would still cut the machine-stiff padding like "in due course" that no register actually needs.
This is where translations give themselves away fastest. Idioms almost never translate word for word. When a tool renders "it costs an arm and a leg" literally from another language, or keeps a source-language saying intact, the result is a phrase that no native speaker uses. A human reader trips on it immediately.
The fix is to find the idea behind the idiom and swap in the natural equivalent your target language already has. Sometimes there is a clean match. Other times the cleanest move is to drop the figure of speech entirely and just say the plain meaning.
Before: "After the merger, the two teams were like cats and dogs in the same bag."
After: "After the merger, the two teams clashed constantly."
The literal version is vivid and wrong. The natural version keeps the meaning, sounds like English, and does not make the reader pause to decode it. When you are unsure whether an idiom landed, that hesitation is your signal to rewrite it.
Proverbs are idioms in formal clothing and they fail the same way. A translated proverb that is famous in the source language often means nothing in the target. Replace it with the closest local saying, or restate the point directly. The reader should never have to guess what a folksy line is gesturing at.
Machine output tends to march. Sentences come out at a similar length with a similar build, and that evenness is exactly the texture human writing does not have. Real prose breathes. A long sentence is followed by a short one. A fragment lands for emphasis. Reading should have a pulse.
Look at your draft as a shape, not just as words. If five sentences in a row are roughly the same length, break one in half and let another run longer. Start a sentence with something other than the subject now and then. The content does not change, but the rhythm starts to feel written rather than generated.
Before: "The product launched in March. The team worked for six months on it. The response from users was positive. The company plans to expand the features. The next release is scheduled for fall."
After: "The product launched in March after six months of work. Users liked it. Now the company is expanding the features, with the next release set for fall."
Same facts, same order, same meaning. The second version varies the length and joins ideas the way a person would, so it stops sounding like a list read aloud.
False friends are words that look the same across two languages but mean different things. "Actual" in several languages means "current," not "real." "Eventual" can mean "possible" rather than "final." Machine translation usually gets these right, but it still leans on the most literal dictionary choice, which can be technically defensible and completely unnatural.
Go through and ask, of each noticeable word, "is this what a native speaker would actually reach for here?" Often the literal term is correct but cold, and a more common word fits better. This is also where you catch the slightly-wrong register words a tool keeps when a simpler one would do.
Before: "We must realize the project before the term expires, according to the disposition of the contract."
After: "We need to finish the project before the deadline, as the contract requires."
"Realize," "term," and "disposition" are all near misses carried over too literally. The natural version trades each for the word an English speaker would use without thinking. The meaning is untouched.
Translationese is wordy. It overuses small connecting words like "the," "of," "in," and "that," and it pads sentences with phrases that add nothing. Part of this comes from the source grammar bleeding through, and part is the tool's preference for safe, complete-sounding constructions. Tightening the text removes the bloat and immediately makes it read more like deliberate human writing.
Hunt for two things. First, function words you can delete without losing meaning. Second, filler phrases like "in order to," "it is important to note that," and "due to the fact that" that almost always have a one-word replacement.
Before: "In order to be able to complete the analysis of the data, it is necessary that we have access to all of the files that are relevant."
After: "To finish the data analysis, we need all the relevant files."
The first version is twenty-seven words of mostly scaffolding. The second is eleven words and says exactly the same thing. Trimming this padding is one of the fastest ways to strip the machine feel out of a passage.
The final pass is your ear, not your eye. Reading aloud forces you through the natural pauses and stresses of speech, and anything that still sounds translated will snag your voice. Where you stumble, slow down, or have to reread to get the sense, that line needs work. This catches the subtle problems the earlier steps missed: an odd word order, a clause in the wrong place, a tone that drifts mid-paragraph.
Before: "Important it is that the report, which was requested by the manager yesterday, be delivered with promptness."
After: "The report the manager asked for yesterday needs to go out quickly."
You can hear the difference the moment you say both out loud. The first fights your voice; the second flows. If you cannot read a passage smoothly, a reader cannot read it smoothly either, so trust the stumble and fix it.
Run the steps in order and most machine translations clean up in one careful pass. Register sets the tone, idiom and false-friend fixes remove the obvious tells, rhythm and trimming handle the texture, and the read-aloud catches the rest. None of it changes what the text means. It changes how a person experiences reading it, which is the only part the machine got wrong.
The catch is time. A paragraph is a five-minute job once you have the rhythm. A long document, a batch of product descriptions, or a thesis chapter is hours of the same close work, and the quality of the edit drops as you tire. That is the practical limit of doing it entirely by hand. The chapter on translationese lays out the full list of the underlying signals you are fixing.
The hand method is the right call for short, high-stakes text where every line matters: a cover letter, an important email, a paragraph you will be judged on. When the volume is large and the steps start to feel mechanical, a humanizer applies the same kinds of fixes at speed.
If your specific problem is a translation that reads as robotic, our focused walkthrough on humanizing translated text covers the same flow with translation examples. And again, if you want the reasons behind every tell you are fixing, why AI translations sound robotic is the explainer that pairs with this how-to.
Sometimes you have a translation already, it reads badly, and you need it to read well. This chapter is that narrower job.
You start with a translation that technically says the right thing but reads wrong. Maybe DeepL or ChatGPT or Google Translate handed you a paragraph where every sentence has the same shape, the idioms landed flat, and the rhythm feels off. You paste that into the humanizer, and it rewrites the passage to sound natural in the target language.
Here is what stays the same and what changes:
You can try it free and see the before and after side by side on the AI Humanizer. Seeing both versions next to each other is the fastest way to understand what the tool actually changed.
Plenty of honest situations leave you holding a robotic translation.
In every one of these, the work is yours and the meaning is yours. The translation step just flattened the voice, and the humanizer puts the voice back.
To fix the flatness, it helps to know where it comes from. Machine translation tools are good at accuracy and bad at sounding human, and that gap has a name and a few clear causes.
Linguists call the telltale flavor of machine output translationese. The grammar is correct, but the writing keeps the shape of the source language instead of flowing the way someone would write from scratch in the target. You can feel it even when you cannot name it. The sentences are technically fine, yet the whole thing reads like a translation. That feeling is the problem the humanizer is solving. We go deeper on the linguistics of it in why AI translations sound robotic.
Every language orders its words and builds its clauses differently. A literal translation tends to drag the source language's structure along with it. Word order that was natural in the original sits awkwardly in the target. Articles and prepositions get used the way the source uses them, not the way the target does. The result is grammatically valid and subtly foreign at the same time.
Translation engines tend to reach for the most statistically likely wording, again and again. So you get the same connectors, the same sentence openings, and a narrow band of vocabulary across a whole passage. Human writing wanders more. We vary sentence length, swap synonyms without thinking about it, and break our own patterns. Machine output does not, and that uniformity is a big part of why it reads as stiff.
Expressions that work in one language rarely survive a literal jump to another. A phrase that is vivid in the source becomes confusing or flat in the target when it is moved word for word. These literal idioms are some of the clearest signs that a human did not write the final draft.
You have two paths, and they work well together: run it through the humanizer, then read it once more by hand.
The fastest fix is to paste the passage into the AI Humanizer and let it rewrite the translationese for you. It varies the rhythm, replaces stiff literal constructions, and smooths the carryover from the source grammar, all while holding your meaning steady. For most translated drafts this gets you most of the way in seconds, and the before-and-after view shows you exactly what moved. The previous chapter has more on the goal of that process.
No tool replaces one careful read. Once you have a cleaner draft, go through it yourself with these moves:
The combination is the point. The tool does the heavy lifting on the patterns that are tedious to fix by hand, and your read adds the judgment a tool cannot, so the final piece sounds like you.
A well-humanized translation passes a few plain tests. The sentences vary in length and shape instead of marching in lockstep. The idioms read like things a native speaker would actually say. Nothing makes you stumble or reread to figure out what was meant. And, most important, every fact and claim from your original is still there, unchanged. If the meaning shifted, the rewrite went too far, and that is not the goal. Good output is your message, said the way a fluent writer would say it.
Here is a small example so the difference is concrete. Picture a sentence that came back word for word from a French source:
Before: It is necessary to take into account that the team, it has finished the project in advance of the foreseen deadline.
After: Keep in mind that the team wrapped up the project ahead of schedule.
What changed: the literal idiom "it is necessary to take into account" became the natural "keep in mind", the doubled subject ("the team, it has finished") was cleaned up, and the stiff "in advance of the foreseen deadline" relaxed into "ahead of schedule." The shorter rhythm reads like a person wrote it. The meaning is identical. The team still finished early, and nothing was added or dropped.
This is also why we show the before and after rather than just handing you a finished block. You can confirm at a glance that the meaning held and only the phrasing improved. If you want to understand the underlying tells the rewrite is smoothing out, the cluster opening chapters on machine translation detection and translationese breaks down every signal in detail.
The best way to see what this does is to run a paragraph of your own through it. Take a translation that reads a little wooden, paste it into the AI Humanizer, and compare the two versions. You will see the repetitive openings break up, the literal phrasing relax, and the rhythm start to sound like writing instead of output. It is free to try, and you keep full control: you read the result, you accept the parts that work, and you keep editing until it sounds like you.
That is the whole promise here. Not a shortcut around anyone, not a trick, just a faster way to turn a mechanical translation into clear, natural writing that still says exactly what you meant.
The honest answer is that it depends on rules you can go and read, and that most people never do. This chapter is about finding out where you actually stand rather than guessing.
Most academic-integrity policies are not written around translation specifically. They are written around two older ideas: the work has to be yours, and you have to be honest about how you produced it. Translation runs into both.
So the useful question is not "is AI translation allowed?" but "am I representing my own ideas, and have I disclosed the tools I used where my institution asks me to?" Translating your own draft from your first language into English, with disclosure where required, sits very differently from translating a published article and submitting the result as original analysis. One is a language tool helping you express your own work. The other is using a different language as a hiding place for someone else's.
Keep that line in mind for the rest of this chapter, because almost every gray area resolves once you ask which side of it you are on.
There are plenty of honest reasons to translate, and on their own they are not misconduct.
What these cases share is that the intellectual work is yours. The tool handled language mechanics, not thinking. Many schools are completely fine with this, and some explicitly allow translation and grammar tools. The catch is that "many" is not "all," and "fine on its own" is not the same as "fine without disclosure." Which is exactly why the policy matters.
The other side of the line is just as clear.
The common thread is misrepresentation: claiming work or ideas that are not yours, or hiding how the work was made when you were asked to be open about it. That is what integrity systems are designed to find, whatever language it happens in.
There is no single global rule here, and that trips a lot of students up. One professor encourages translation tools for multilingual students as an accessibility aid. Another in the same department treats any machine translation as unauthorized assistance. A journal may allow it with a method note; a scholarship competition may ban it outright.
This variation is real, and it is the reason you cannot rely on what a friend at another school was told. Policies differ by institution, by department, sometimes by individual instructor and individual assignment. A language-learning course will treat translation very differently from a literature seminar.
Disclosure is what cuts through the uncertainty. When a policy is unclear or silent, a short, honest note about what you did and which tool you used almost always protects you, because the thing integrity policies punish is concealment, not honest help. "I drafted this in Spanish and translated it to English using DeepL, then revised it myself" is a sentence that turns a possible problem into a non-issue. If you are unsure whether to disclose, disclose. It is the cheapest insurance there is.
It helps to know what is really being looked at, because the picture is less mysterious than rumor suggests. Turnitin and similar tools work in two separate layers, and a human grader adds a third.
Turnitin can translate non-English submissions into English and compare them against its database of papers, journals, and web pages. This is cross-language matching, and it is the reason translating a source does not produce "original" text. The words changed, but the structure, the order of ideas, and the specific claims usually did not, and that is what the comparison catches. If your paragraph is a translated copy of a published article, the match still tends to surface. We go deeper into this in Can Turnitin detect translated text? and the specific case in Can Turnitin detect Google Translate?.
Separately, Turnitin looks for the statistical fingerprint of machine-generated text. Tools like ChatGPT and DeepL tend to choose very even, predictable phrasing, and translated output often carries that same smoothness. So a translated passage can get flagged by the AI check even when no source match exists, simply because it reads like machine output.
A grader who knows your earlier work and reads carefully is often the most accurate check of all. Translationese, which is text that keeps the grammar and rhythm of the source language, reads oddly to a native ear: literal idioms, repetitive sentence length, a tone that shifts without reason. A teacher who sees a sudden jump in fluency, or phrasing that does not match your usual voice, may simply ask you about it. Our machine translation detection opening chapters breaks down those signals in full.
Knowing these three layers is not about gaming them. It is about understanding that translation does not erase the trail, so the honest path is also the practical one.
Here is the part that deserves the most care, because it cuts the other way and it matters for fairness.
Detection is probabilistic, not a verdict. A score is a confidence estimate. A "40% AI" result does not mean 40 percent of your essay was written by a machine. It means the model is partly unsure. And these models make mistakes in both directions.
The mistake that should worry multilingual students is the false positive. Writing in a second language can look "too even" in exactly the way machine text does: careful word choices, steadier sentence rhythm, fewer of the messy quirks a first-language writer leaves behind. Independent reviews have found that detectors flag non-native English writers more often than native writers, and that real-world detection accuracy sits lower than the marketing suggests, dropping further on translated content (testing has put the accuracy loss on translated text at roughly 20 percent). Some universities have turned their AI detection features off over precisely these fairness concerns.
So if English is not your first language, you can write your own honest essay and still see a detector light up. That is a flaw in the tool, not proof of anything about you. The right response is not to panic and not to start gaming scores. It is to keep the evidence that the work is yours, which is the next section.
None of this requires tricks. It requires a few honest habits.
Notice that every one of these steps makes your work more honest and easier to defend, not harder to catch. That is the point. The goal is not to slip past a detector. It is to be able to stand behind what you submitted.
The most useful move is to see what a detector sees before anyone else does. Running your essay through an AI detector first shows you which passages read as machine-translated or AI-smoothed, so you can revise the weak spots in your own voice rather than find out after the fact.
The uncomfortable thing about this subject is that the people most likely to be wrongly flagged are the people with the least power to argue about it. An international student on a visa, a researcher publishing in their third language, a freelancer in another timezone. Careful English reads as predictable English, and predictable is what these systems were built to punish.
None of that is a reason to write worse. It is a reason to know what the tools are measuring, to keep the evidence of your own process, and to be able to explain in plain terms why a score is not proof. That is most of what this book has been for.
Use the translators. They are extraordinary, and pretending otherwise helps nobody. Just do not hand in the raw output, and do not let a machine decide how your sentences sound. The retelling step in the last two chapters is slower than pasting, and it is the difference between a text that is yours and a text that merely passed.
If it was useful, send it to one person who needs it. There are two companion books, Sounds Like You on keeping your own voice when you draft with AI, and Nobody Wrote That Paper on the citations these tools invent.
You are welcome to share this book with anyone who might find it useful, as long as you share it whole and unchanged and do not sell it. Writing centres, international student services and language departments: please link or hand it out freely, no permission needed.
Copyright © 2026 Dipak Bhosale. Published by Lacewing Technologies, Navi Mumbai, India.
This book is for general information and is not legal or academic advice. Policies on translation and AI use differ by institution and by course, so read your own before you rely on anything here. TextSight is an AI-detection and writing-trust tool. No detector, ours included, can prove that a piece of writing was or was not translated, and accuracy on translated text is lower than on text written directly in English. Treat every score as one probabilistic data point, never as proof.
This book is clear that detectors are less reliable on translated text and that no score proves anything. What a scan is good for is showing which of your sentences carry the machine signal, so you know which ones to retell in your own words before you hand the work in.
