Does Turnitin detect QuillBot? Yes, and it says so by name
Yes, usually, and Turnitin names QuillBot in its own product documentation as an example of what it is built to catch. Since July 2024 its AI writing report has split the score into two categories, one of which is text that was likely AI-generated and then likely revised using an AI-paraphrasing tool or AI word spinner, such as Quillbot. What the detector finds is the AI fingerprint underneath the rewrite rather than QuillBot itself, which is why Turnitin says the same feature does not flag human-written text put through a spinner. The honest caveat is that nobody has published a real test of today's Turnitin against QuillBot, and every precise-sounding percentage circulating on this question comes from a company selling a humanizer. None of that changes the part which survives regardless: most universities treat running text through a paraphrasing tool as misconduct whether or not a detector catches it.
If you are searching this, you have probably already run something through QuillBot and you want to know what happens when you hit submit. The answer comes from the two companies involved rather than from anyone selling a way around them. Turnitin publishes what its detector looks for, QuillBot publishes what its modes do, and setting those two documents side by side dissolves most of the confusion here.
Turnitin names QuillBot in its own documentation
Most articles on this question guess. Turnitin does not make you guess, because it prints the name.
Since a release dated 16 July 2024, the AI writing report has split its overall score into two categories. The first is text likely generated by a large language model. The second, in Turnitin's own words, is qualifying text that was likely AI-generated and then was likely revised using an AI-paraphrasing tool or AI word spinner, such as Quillbot. The same product name appears again in the guide instructors read when they open a report, which describes detection of text further modified by an AI bypasser, AI-paraphrasing tool or AI word spinner, such as Quillbot.
The capability is older than the label. Turnitin's release notes record that on 6 December 2023 it began detecting likely AI-generated text even when paraphrased by an AI word spinner, and that the check runs automatically with no setting an instructor can change. Detection aimed at AI bypassers folded into the same report on 27 August 2025, with further updates in October 2025, February 2026 and July 2026, each described as improving recall.
One oddity is worth noticing. Turnitin's FAQ says it will not publish the list of paraphrasing tools it detects, because doing so would make it easier for students to evade our system. It then names QuillBot twice anyway. Read that as a company comfortable enough with this case to make an example of it.
What the detector is actually looking at
It helps to know that Turnitin is not looking for QuillBot. There is no QuillBot signature and no library of paraphraser output to match against.
Turnitin describes the process plainly. Sentences are extracted from your submission and segmented into overlapping sections. Each segment is scored between 0 and 1 for the probability of being human or AI-generated, each qualifying sentence inherits its segment's score, and because the segments overlap, sentences carrying several scores have them pooled into one. Those scores aggregate into the document percentage an instructor sees, which we unpack further in what your Turnitin AI score means.
What survives a paraphrase, then, is not wording. It is the shape underneath: the order the argument arrives in, how evenly the sentences are weighted, the smoothness that generated prose tends to have. Swapping vocabulary changes the surface and leaves that structure where it was.
That matters if you did the writing yourself. UNSW Sydney, briefing staff on the July 2024 update, records that Turnitin say that the AI paraphrasing detection will not identify originally human-written text that was passed through a text spinner, adding its own reading: the detector is finding AI fingerprints left over from the original generation. If the text was yours before you paraphrased it, this feature is not aimed at you. Your exposure is the similarity score and your university's rules, which are the next two sections.
What QuillBot does to a sentence
QuillBot documents its own behaviour, which spares us the guesswork here too. It is an AI rewriter rather than a thesaurus on a loop, and its help centre lists ten modes.
Two are free. Standard balances rewriting with maintaining meaning, while Fluency fixes grammar and, in QuillBot's phrasing, makes the fewest changes with synonyms minimised. The other eight sit behind Premium. A synonym control adjusts how aggressively words are replaced, and a Rephrase button generates another variation on demand.
One mode is worth reading closely. QuillBot's description of its Humanizer mode says it makes text sound less robotic and more natural, and that it is especially helpful for AI-generated drafts. The company also sells a standalone AI Humanizer, trained to remove AI tells, with a browser extension that does it inside the chatbot, and an AI detector on a neighbouring page. Whether any of that category works is the subject of do AI humanizers actually work.
None of that is hidden, and QuillBot is not coy about the ethics either. Its humanizer page tells you to follow your institution's AI guidelines and states that trying to pass off someone else's idea as your own, even if it has been reworded, is just as unethical as directly copying a quote. That is the vendor's own sentence, and it is a better summary of the situation than most of what ranks for this query.
The two scores move in opposite directions
Almost every page on this topic writes as though Turnitin produces one number. It produces two, and paraphrasing moves them in opposite directions. Turnitin says the similarity score and the AI writing percentage are completely independent and do not influence each other. If that distinction is new, we untangle it in can Turnitin detect ChatGPT.
The similarity score is text matching, and paraphrasing is very good at defeating it. In a 2017 experiment in the International Journal for Educational Integrity, Rogerson and McCarthy took a paragraph from a paper they had published themselves, ran it through two free paraphrasing tools, and submitted all three versions. The original came back at 100% similarity and both rewrites came back at 0%. That was one paragraph, two tools, and not QuillBot, so treat it as a demonstration of a mechanism rather than a measurement of anything current. The mechanism is durable: change the strings and string matching stops finding them.
The AI score can move the other way entirely. Among the things Turnitin lists as producing false positives is text that has been paraphrased without developing new ideas. Flat, low-variation, reworded prose is what the model associates with machine output, whoever produced it, which is part of why honest writers get caught at all, as we cover in will AI detectors flag my writing. A paraphrase can pull the similarity number down while pushing the AI number up. That is a bad trade, and the opposite of what the pages selling a rewriter imply.
The one real test, and the numbers everyone quotes
Here is the part no competing page will tell you. There is almost no current evidence on this question, and the confident percentages you have been reading are not measurements.
Start with what exists. In 2023, Weber-Wulff and seven colleagues published a peer-reviewed test of fourteen detection tools in the International Journal for Educational Integrity. It is the only credible published study that put QuillBot output in front of Turnitin, and its methods are fully described, which is the whole reason to trust it. Nine researchers each produced one document per test case, giving nine per case and 54 in total. One case was ChatGPT output rewritten automatically with the AI-based tool Quillbot, using the default values of the tool for modes (Standard) and synonym level.
Turnitin came first of all fourteen tools at 76% overall, and flagged none of the human-written documents, a clean zero on false positives. On raw ChatGPT text it scored 8 of 9 and 9 of 9. On the QuillBot-paraphrased documents it scored 2 of 9, eight months before it shipped paraphrase detection at all. Across all fourteen tools the average on that case was 15%, against 94% on human writing, and the authors concluded that using AI to transform AI-generated text produces text the classifiers consider human-written.
That 2 of 9 is the only real number on this page, and it is out of date, which you have to hold in the same hand. Turnitin was tested between 14 and 20 April 2023. Paraphrase detection arrived on 6 December 2023, bypasser detection in August 2025, with three recall updates since. The study measured a version of Turnitin that no longer exists, before the feature this article is about had been built, and nobody has published an equivalent test of the current one. That is the honest state of the evidence, and any page telling you otherwise is about to show you a number it invented.
The numbers circulating on this question
We went through the pages ranking for this query. Nearly all sell an AI humanizer, and every single one says Turnitin catches QuillBot. That is not a consensus of evidence but a consensus of incentive, because "Turnitin catches QuillBot" is precisely the claim that sells a humanizer subscription. Precision and evidence then run in opposite directions: the one page we found showing screenshots of a real Turnitin submission quotes no percentage at all, while the pages quoting figures to the decimal place describe no method whatsoever.
Take the most detailed of them, which reports a controlled test of 50 texts in February 2026 and helpfully prints its subgroup counts next to its percentages. The counts and the percentages do not agree, and not once:
| The count it states | The percentage it reports | What that count actually is |
|---|---|---|
| 14 of 20 | 68% | 70% |
| 10 of 15 | 65% | 66.7% |
| 8 of 15 | 55% | 53.3% |
| 32 of 50 | 62% | 64% |
Four subgroups, four disagreements, none of them rounding. A page reporting real measurements cannot fail its own division four times out of four. Those percentages were not calculated from those counts, because there were no counts. The same page advertises a bypass rate above 99% for its own product.
The pattern repeats. One vendor publishes ten sample scores spanning 64% to 99% from ten texts it says it generated itself, with no screenshots, dates or account details; two other sites reprint that range as a finding without naming where it came from. Several publish a ladder of per-mode scores implying some QuillBot settings fare better than others. None of those ladders agree with each other, which is what you would expect if they were written rather than measured, and it is why we are not reproducing them.
Then there are the studies. Search this question and you will meet a 2024 study in the Journal of Educational Technology, a 2024 study in the Journal of Academic Ethics, and a 2023 University of Toronto study, each with a tidy percentage attached. We searched CrossRef for all three and found none of them. A general search for QuillBot and Turnitin detection returns mostly papers on how language students feel about QuillBot as a learning aid, not detection experiments at all. Until somebody produces a DOI, treat those three as furniture.
Finally, the Turnitin boilerplate, the most mangled item on the query. A 98% accuracy figure appears on page after page, attributed to Turnitin. Turnitin's current FAQ does not contain it. The false-positive rate below 1% is real, but applies only to documents scored above 20% AI, a condition dropped every time it is quoted. Turnitin also says it accepts missing some AI writing to hold that rate down: if it identifies 50% of a document as likely AI, the document could contain as much as 65%. A company telling you it undercounts on purpose is not one claiming to catch everything.
What your university already says
Everything above is a question about a tool. Your institution has already answered a more durable question about conduct, and its answer does not move when the tool does.
Liberty University's faculty guidance has a section headed Paraphrasing Tools. It names QuillBot, notes its output is highly likely to be flagged, then says the part that matters: the university's standards do not permit the use of paraphrasing tools to the extent that originality becomes compromised, and letting your work be heavily manipulated by one means it is no longer in your own words, which is considered academic misconduct. Western Sydney University answers the question students actually ask, whether you can use QuillBot on a source you found online, with in short, the answer is no. The University of Lancashire's 2025 policy goes further still, listing attempting to use AI to humanise or hide the use of AI as an offence in its own right, which makes the concealment the misconduct regardless of whether it worked.
Curtin stopped using Turnitin's AI detector in 2026. The rule did not move. Curtin's student guild, quoting the university's own academic integrity guide, tells students that as of 2026 the University is no longer using Turnitin's AI detector, however it is still forbidden to use AI without permission, and on paraphrasing specifically: if you put the words in one order, and an AI puts them in a different order, you are committing academic misconduct. The detector was the thing you were trying to beat. It was never the thing that decides.
Institutions differ at the edges. Some are relaxed about a grammar pass over sentences you wrote yourself; Curtin and Liberty are not. Where they converge is the case this article is about: taking text you did not write, from a source or from a chatbot, running it through a paraphraser and submitting the result. Read your own course's policy, and our guide to whether using AI to write essays counts as cheating covers how these rules are usually drawn.
The version with nothing to launder
Step back and notice what the question assumes. You have text you did not write, you would rather nobody knew, and you are pricing the odds. Every risk in this article lives downstream of that one arrangement.
The alternative is not working harder. It is not producing text that needs disguising. Read real sources, keep them attached to the claims they support, and write the argument yourself with help rather than instead of yourself. Done that way the detector question stops being interesting, because there is nothing underneath the prose a rewrite would need to hide.
That is what CiteOwl is built for, and it is deliberately not a humanizer. There is no undetectable mode and we would not ship one. It is an agent that researches real papers, links every factual claim to a source you can open, and proposes every change as a diff you accept, reject or edit, so the sentences that land are ones you chose. Working from real sources with real decisions on every line is simply a better way to write than laundering something you cannot defend. The paper comes out yours, and the question you arrived with never comes up.