Sometimes — and the answer depends entirely on which kind of paraphrasing you mean. Lightly word-swapped source text is often still caught by the Similarity Report. AI-paraphrased AI text frequently still reads as machine-generated to the AI indicator. Genuine rewriting in your own words and structure typically isn't flagged by either system — because at that point, it's just writing.
"Can Turnitin detect paraphrasing" is really three questions wearing one trench coat, and most articles on this topic fail by answering only one of them while letting you believe they've answered yours. So let's separate them properly. When people type that search, they mean one of three things: rewording a source — a journal article, a Wikipedia page — for an essay; rewording AI-generated text so it reads less like AI; or running text through a paraphrasing tool like QuillBot and hoping the machine's rewording counts as their own. Different mechanisms are watching each one, the evidence for each is different, and — fair warning — this article contains no bypass tricks for any of them, because we don't publish those and the honest analysis explains why they wouldn't be worth much anyway.
One piece of architecture first, because everything below depends on it. Turnitin is two systems in one report. The Similarity Report — the original product, two decades old — compares your text against a massive database of student papers, web pages, and publications, looking for matching and near-matching strings. The AI writing indicator, added April 4, 2023, is a statistical classifier estimating whether prose was machine-generated, using measures like perplexity and burstiness (unpacked in our how detectors work explainer). Paraphrasing interacts with these two systems in almost opposite ways. Keep them straight and the whole topic gets simpler.
Case one: paraphrasing a source — the similarity side
You read a paragraph in a journal article, you rephrase it for your essay, you cite it (or you don't — hold that thought). Does Turnitin notice?
The Similarity Report is fundamentally a matching engine. Exact quotes and near-verbatim strings light up reliably. Turnitin's matching also has some tolerance for small perturbations — a swapped word here, a reordered clause there — so the classic freshman move of replacing every fourth word with a thesaurus synonym frequently still matches, because enough of the underlying string survives. What patchwriters consistently underestimate is how much of a sentence stays intact when you only swap the "big" words: the connective tissue, the phrase order, the distinctive number-and-name sequences all still align with the source.
Push the rewording far enough, though, and the honest answer is that string matching runs out. A paragraph fully reconstructed — new sentence boundaries, new order of ideas, your vocabulary throughout — typically won't match the source in the Similarity Report at all. Turnitin has been candid over the years that similarity checking finds matching text, not unattributed ideas; the report literally measures overlap, and a skilled paraphrase has little.
Here's where students draw exactly the wrong conclusion. A low similarity score does not mean you're in the clear — it means the software has nothing to say. Paraphrasing a source without citation is plagiarism whether or not any string matches; you've taken someone's ideas and presented them as yours. And the detector for that isn't Turnitin. It's your instructor, who assigned the same readings, recognizes the argument of the article you reworded, and notices that your essay reproduces its structure claim by claim with no citation in sight. Mosaic plagiarism cases get opened by humans far more often than by software. The fix costs one parenthetical: cite the source, and paraphrasing stops being a risk at all — it becomes the skill your instructors actually want, since properly attributed paraphrase is the backbone of academic writing.
Case two: AI text, paraphrased by AI
Now the version of the question that spiked after 2023: you generated an essay with ChatGPT or Claude, and you ran it through another AI pass — "reword this so it sounds human," or a second model paraphrasing the first — hoping to drop the AI score. Does that work?
Give the idea its due first, because it isn't baseless. Academic research on how well detectors withstand tampering has repeatedly shown that paraphrasing attacks can degrade AI-text detectors — rewording machine text was one of the earliest demonstrated weaknesses in the detection literature. Weber-Wulff et al. measured it directly: across 14 detectors, accuracy on machine-paraphrased AI text was 26%. Detection is adversarial territory, and anyone claiming detectors are immune to paraphrase is overclaiming in the other direction.
But three facts should temper the enthusiasm considerably. First, an AI paraphrase of AI text is still AI text. The paraphrasing model selects high-probability words in evenly-shaped sentences, because that's what language models do — you've laundered the wording while regenerating the very statistics the classifier measures. Machine-reworded machine prose routinely still reads machine-typical. Second, the vendors moved. Turnitin has announced that its detection extends toward AI-paraphrased and humanizer-processed text, with AI paraphrasing detection added in July 2024 and AI bypasser detection in August 2025, both English-only and neither carrying published accuracy figures. We treat vendor claims with the same skepticism we apply to detector accuracy claims generally — Turnitin's headline 98% figure only ever applied to documents flagged above 20%, and its own interface replaces 1–19% scores with an asterisk precisely because low scores are unreliable. But the direction is clear: the paraphrase loophole is a moving target that the largest player in the market is actively paying engineers to close. Third, even where the trick works today, you've built your academic standing on an arms race you don't control. Detectors retrain; your submitted paper is in the database forever; reprocessing of archived work is a policy decision away.
And beneath all of it sits the fact no statistical argument touches: if AI writing is banned in your course, AI writing that's been AI-paraphrased is still banned, and disguising it is an integrity violation on its own — often a worse one, since it demonstrates intent. Detection odds don't change what the rules say. We've said this in every post in this series and it stays true in this one.
Case three: paraphrasing tools — QuillBot and its cousins
The third meaning is the most searched: dedicated paraphrasing tools. QuillBot is the household name; there are dozens. Students use them on sources (case one with automation) and on AI output (case two with a cheaper engine), so both analyses above apply — but the tools deserve their own entry because their output has its own signature.
Paraphrasing tools are typically smaller rewriting models. They preserve sentence skeletons while substituting synonyms and shuffling phrases, which produces two detectable artifacts. Against the Similarity Report, the preserved skeleton means source structure often survives well enough to partially match — word-spinner output is famously "the same sentence in a cheap disguise." Against the AI indicator, tool output is machine-generated text, full stop, and frequently scores as such. There's also a tell no software is needed for: synonym substitution without understanding produces the genre's distinctive wrongness — "key factors" becoming "pivotal elements," idioms mangled, terms of art swapped for near-synonyms that no one in the field would use. Instructors have been recognizing spun text since long before ChatGPT; article-spinning for SEO trained a generation of readers to spot it.
Turnitin's expansion toward paraphrase-and-humanizer detection is aimed squarely at this category. We've written a full test-and-evidence breakdown in Can Turnitin detect QuillBot?, and the adjacent question — AI text processed by tools that market themselves as "humanizers" — gets its own treatment in Does Turnitin detect humanized AI?. The one-line preview of both: tool-processed text is caught inconsistently, which is precisely why no honest company promises otherwise.
The three cases side by side
| What you actually did | Similarity Report | AI writing indicator | Realistic risk |
|---|---|---|---|
| Quoted or lightly word-swapped a source | Often matches — strings largely survive | Not the relevant system | High, and it looks deliberate |
| Deeply rewrote a source, with citation | Little to match | Human-written prose; typically not flagged | Low — this is proper academic writing |
| Deeply rewrote a source, no citation | Little to match | Typically not flagged | Software-low, human-high: plagiarism if recognized |
| AI-generated text, submitted raw | N/A (usually novel text) | Commonly flagged; the case detectors handle best | High |
| AI text paraphrased by AI or a tool | Sometimes partial matches | Explicitly targeted since July 2024 — but 26% detector accuracy on paraphrased AI text in peer-reviewed testing | Unstable — and rising as detectors retrain |
| AI draft rebuilt in your own words, ideas, structure (where permitted) | Little to match | Increasingly reads as human, because it increasingly is | Low on detection; policy compliance is the real question |
The table's diagonal tells the story. Every low-risk row involves you doing the intellectual work. Every high-risk row involves outsourcing it and hoping software can't tell.
The honest middle: rewriting versus word-swapping
Strip away the tools and the acronyms, and the paraphrasing question reduces to something older than Turnitin: the difference between reprocessing text and re-expressing ideas.
Word-swapping — by thesaurus, by QuillBot, by "reword this" prompt — leaves the underlying composition intact. Someone else (or something else) chose the argument, the order, the emphasis, the evidence; you changed the paint. Both of Turnitin's systems are, in their different ways, decent at noticing paint jobs, and both are getting better at exactly this. More to the point, so is any instructor who reads carefully, and they don't have a false-positive asterisk policy.
Deep rewriting is a different act. You understood the material, closed the tab, and rebuilt the point inside your own argument — your structure, your examples, your voice, with the source cited. That text doesn't evade detectors; it never triggers them, because nothing about it is machine-typical or copied. This isn't a loophole. It's the assignment. The entire reason paraphrase-with-citation is taught as a core skill is that re-expressing an idea accurately is how you prove you understood it.
Where does an AI humanizer sit in this picture? Exactly where we always say it does. HumanFlow's AI humanizer rewrites text toward more natural, varied prose — and it does not promise to beat Turnitin or any detector, because nobody can honestly promise that, and this article is a 3,000-word explanation of why. Our position, published and repeated: the legitimate use is improving text you're allowed to use AI on and intend to own — and then editing the output into something that's genuinely yours — not disguising banned work. If a tool in this category does promise you an undetectable result, you now know enough to price that promise correctly.
What this means in practice
For the student paraphrasing sources: cite everything you reword, aim for full re-expression rather than synonym swaps, and the Similarity Report becomes irrelevant to you — a well-cited paraphrase with 2% similarity is a normal, good outcome. For the student tempted by the paraphrase-the-AI route: the statistics are unreliable in your favor, the vendor is actively closing the gap, and the rule you'd be evading applies regardless of the score. For instructors: a low similarity score was never proof of originality, an asterisked AI score was never proof of anything, and the reading you do yourself remains the best detector in the building — Turnitin's own design choices, like refusing to print scores under 20%, quietly agree.
For everyone: the pillar guide to Turnitin AI detection covers the thresholds, the false-positive record (including the Patterns study that found detectors falsely flagging 61% of non-native speakers' essays on average), and what the scores do and don't prove.
FAQ
Can Turnitin detect paraphrasing of a source if I change most of the words? The Similarity Report matches text strings, so a paraphrase that genuinely reconstructs the sentences usually won't match. That protects you from a similarity flag, not from plagiarism: an uncited paraphrase is still plagiarism, and instructors recognize reworded arguments from readings they assigned. Cite it and the question disappears.
Does paraphrasing AI-generated text remove the AI score? Unreliably at best. AI-paraphrased AI text is still machine-generated at the statistical level detectors measure, and Turnitin says its detection now extends to paraphrased and humanizer-processed text. Research shows paraphrase attacks can degrade detectors, but the effect is inconsistent and shrinking as systems retrain.
Can Turnitin detect QuillBot specifically? Turnitin doesn't name tools, but spun text is exposed on two fronts: preserved sentence skeletons can still partially match sources, and tool output is itself machine-generated text the AI indicator can flag. Our QuillBot deep-dive covers the evidence in detail.
Is paraphrasing with citation ever a problem? Almost never, when the re-expression is genuine. Cited paraphrase is a core academic skill, and a low similarity score on well-cited paraphrase is the intended outcome. The gray zone is "patchwriting" — cited but too close to the original's wording — which many instructors treat as a citation-quality problem rather than misconduct, though policies vary.
What similarity score counts as paraphrasing plagiarism? No score does, in either direction. Similarity measures overlap, not intent or attribution: a 40% score can be fine (heavily quoted, fully cited) and a 3% score can hide wholesale uncited paraphrase. Any instructor treating the number itself as a verdict is misusing the tool, and Turnitin's own guidance says the report requires human interpretation.
If my rewritten AI draft passes detection, is it my work? Detection and authorship are separate questions. If your course permits AI assistance and you rebuilt the draft with your own argument and phrasing, disclose per policy and you're on solid ground. If AI use was banned, a passing score changes nothing — the work still violates the rule, and version history or an oral defense can surface that later.
Did Turnitin really add paraphrase and humanizer detection? Yes — AI paraphrasing detection in July 2024 and AI bypasser detection in August 2025, both English-only. Independent verification of how well it works is thin so far, which is reason for skepticism about the marketing in both directions — neither "it catches everything" nor "it catches nothing" is supported.
Key facts
- Turnitin runs two systems: the Similarity Report (text matching, ~two decades old) and the AI writing indicator (statistical classifier, launched April 4, 2023) — paraphrasing interacts differently with each (Turnitin).
- Turnitin's 98% accuracy / <1% false-positive claims apply only to documents flagged above 20% AI; scores of 1–19% display as an asterisk because Turnitin considers them unreliable (Turnitin AI writing FAQ).
- Turnitin screened 200M+ papers in year one; ~11% showed ≥20% AI writing, ~3% were ≥80% AI (Turnitin, April 2024).
- Turnitin added AI paraphrasing detection (July 2024) and AI bypasser detection (August 2025), both English-only, with no published accuracy figures for either (Turnitin).
- Liang et al. (Patterns, 2023) found seven AI detectors falsely flagged an average of 61.22% of 91 human-written TOEFL essays — a caution against treating any flag, including a paraphrase flag, as proof.
- OpenAI retired its own AI text classifier in July 2023 after it identified only 26% of AI text (OpenAI) — the benchmark for humility in this field.
- Weber-Wulff et al. (2023): across 14 detectors, accuracy on machine-paraphrased AI text was 26% — against below-80% overall accuracy for every tool tested.
Sources
- Turnitin — AI writing detection FAQ and transparency page (system architecture, thresholds, asterisk policy).
- Turnitin — press releases on AI paraphrasing detection (July 16, 2024) and AI bypasser detection (August 27, 2025).
- Turnitin — first-anniversary press release, April 2024 (200M+ papers; prevalence figures).
- Liang, W. et al., "GPT detectors are biased against non-native English writers," Patterns (Cell Press), 2023.
- OpenAI — announcement retiring its AI text classifier, July 2023.
- Krishna, K., Song, Y., Karpinska, M., Wieting, J., & Iyyer, M. — "Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense." NeurIPS 36, 2023.
- Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. — "Can AI-Generated Text be Reliably Detected?" Transactions on Machine Learning Research.
- Perkins, M., Roe, J., et al. — "Simple techniques to bypass GenAI text detectors: implications for inclusive education." International Journal of Educational Technology in Higher Education, 21:53, 2024.
- Turnitin — Similarity Report interpretation guidance (similarity vs. plagiarism distinction).