Does Turnitin Detect QuillBot? What 2026 Testing Shows
Running an AI draft through QuillBot is the most common workaround students try, and it is the one with the worst risk-to-benefit ratio. Turnitin has been trained on paraphraser output since late 2023, it flags paraphrased AI in a separate category from generated AI, and independent 2026 tests put its catch rate on QuillBot output near 70 percent. Here is what the numbers actually say, why synonym swapping does not work, and what does reduce risk.
The short answer: yes, in roughly 7 out of 10 cases
Turnitin flags most QuillBot-paraphrased AI text. Across the independent tests published through mid-2026, the average detection rate on QuillBot output sits near 70 percent, and Turnitin's own model documentation describes a wide 64 to 99 percent range depending on mode, text length and the model that wrote the original. One controlled February 2026 run of 50 ChatGPT-4o texts through QuillBot Standard reported 62 percent flagged; a separate June 2026 sample found Turnitin caught every Standard-mode essay it was given.
The spread matters more than the average. Detection climbs with length, because a longer text gives the classifier more data points: the same 2026 test flagged 68 percent of 800 to 1,200 word argumentative essays but only 55 percent of 150 to 300 word short answers. No point in that range is a safe bet. The variance means your individual outcome is a coin flip whose odds you do not control and cannot check before you submit.
| QuillBot mode | Reported Turnitin detection rate | Typical flag category |
|---|---|---|
| Standard (free) | ~80% | AI-paraphrased |
| Fluency (free) | ~75% | AI-paraphrased |
| Formal (paid) | ~72% | AI-paraphrased |
| Creative (paid) | ~55% | Mixed |
| Custom, maximum changes | ~50% | Mixed |
Mode-level figures are third-party estimates published in 2026, not Turnitin's own per-mode disclosures. Treat them as directional, not as a threshold to optimise against.
Why swapping words does not remove the AI signal
QuillBot is a paraphraser, not a detection tool. It substitutes synonyms, reorders clauses and shifts register between modes. What it does not do is change the property detectors measure. Turnitin does not compare your vocabulary against a database; it models how predictable a sequence of words is, which the field calls perplexity, and how much sentence length and complexity vary, which it calls burstiness. Standard mode rewrites the surface while leaving paragraph rhythm and uniform sentence lengths intact, and that shape is precisely what the classifier was trained to recognise.
There is a second reason the workaround fails: paraphrased text is not an unmodelled edge case for Turnitin, it is a labelled training category. The AIW-2 model released in December 2023 added dedicated paraphrase detection, the AIR-1 model in July 2024 extended it to rewritten text, and a 2025 update trained across the outputs of multiple paraphrasing and humanizing tools at once. The paraphrasing step itself leaves a recognisable distortion pattern, because a paraphraser also follows probability rules when it picks a synonym.
Passing Turnitin is not the same as passing detection
Turnitin is the most forgiving of the mainstream detectors on paraphrased text, and that is a deliberate design choice: it accepts lower recall in order to hold its false-positive rate near 1 percent. It would rather miss paraphrased AI than accuse a student who wrote their own work. That trade-off is the only reason QuillBot output clears it as often as it does.
The problem is that the lenient tool is the one your institution licenses, not the only one you will face. Originality.ai publishes roughly 94.66 percent accuracy against QuillBot-paraphrased content specifically, because it was trained on paraphrased AI as a primary target. Independent comparisons put GPTZero near 91 percent and Copyleaks in the 75 to 85 percent band on the same material. An instructor who pastes a suspicious passage into a second tool, which a growing number now do, sees a far higher score than the one in the similarity report.
The statistical tell, in plain terms
Human writing is uneven. We start a sentence one way, change direction halfway through, drop in an odd word choice, then follow a long winding clause with a three-word fragment. That unevenness is what burstiness measures, and it is the single hardest thing for a paraphraser to manufacture, because a paraphraser works one sentence at a time and has no reason to vary the shape of the paragraph.
This is why GPTZero reads QuillBot output so clearly at around 91 percent. QuillBot produces text that is smooth, predictably structured and uniform in sentence length, which is low perplexity and low burstiness at the same time, the opposite of the human baseline. It is also worth knowing that GPTZero's free tier catches paraphrased text at close to the same rate as its paid tier; paying more for a detector does not meaningfully change your exposure, because the detection model is the same.
One weak detector clearing you means nothing
In a June 2026 comparison across five detectors, Sapling was the clear outlier: it flagged 3 of 5 raw AI texts, 1 of 5 QuillBot Standard rewrites and none of the Creative-mode rewrites. Independent benchmarks put it around 90 percent at best on clean AI output, lower on edited text, and it is a general-purpose NLP suite where detection sits alongside grammar and sentiment endpoints rather than being the core product.
That makes it the wrong instrument for a self-check. Testing against the most permissive tool you can find and calling the result undetectable is selection bias, not verification. If you are going to check anything, check against the tool that will actually judge the work, and treat a clean result as the absence of evidence rather than evidence of safety.
The flag that actually costs you: paraphrased, not generated
Turnitin's instructor dashboard separates two outcomes with different colours. Cyan marks text identified as AI-generated. Purple marks text identified as AI-generated and then paraphrased. The second label is the more damaging one, because it does not just suggest AI was used, it suggests a step was taken to disguise it, and academic integrity processes routinely treat concealment more harshly than the underlying conduct. Paraphrasing is therefore the worst of both worlds: it keeps most of the detection risk and adds an inference of intent.
What genuinely lowers risk is editing at the level of substance rather than vocabulary. For calibration, Grammarly's rewrite feature is even weaker than QuillBot here: it targets grammar and clarity, so 85 to 95 percent of the AI signal survives it, and manual synonym swapping by hand performs no better at 75 to 85 percent detection. What changes the statistical shape of a text is adding your own argument, your own evidence, your own examples and your own errors. That version is also the only one that survives a five-minute conversation with the person grading it.
Bottom line
Turnitin catches QuillBot-paraphrased AI text roughly 70 percent of the time, more on longer submissions, and it labels the result as AI-paraphrased rather than AI-generated, which reads as an attempt to conceal. It is also the most lenient mainstream detector on paraphrased text, so clearing it says little about Originality.ai at about 95 percent or GPTZero at about 91 percent. Paraphrasing keeps the risk and adds the appearance of intent. Rewriting with your own reasoning and evidence is the only approach that changes the statistics and holds up to a follow-up question.