Can paraphrase tools be detected?
Short answer
Partly. Turnitin's AI writing FAQ says it detects AI text modified by paraphraser and bypasser tools, for English submissions only. Research says caution: in Krishna et al. (2023), paraphrasing cut DetectGPT's accuracy from 70.3% to 4.6% at a 1% false positive rate. A low score on paraphrased text proves little.
No detector score on paraphrased text should be the only evidence in a misconduct case.
"Paraphrasing" covers two different problems. One is a student running ChatGPT output through a rewriting tool so an AI detector misses it. The other is a student rewording someone else's source so a plagiarism checker misses it. Detectors are getting better at the first, and plagiarism checkers have long worked on the second, but neither problem is solved. This guide covers both, from a teacher's side of the desk.
What each detector documents about paraphrased AI text
| Detector | Claims to flag paraphrased AI text? | Published number | Main limit |
|---|---|---|---|
| Turnitin | Yes, built into the AI writing indicator | Under 1% false positives for documents over 20% AI (overall model) | English only; tools not named |
| GPTZero | Yes, in beta | About 95% on certain bypassers; under 0.1% false positives | Covers "certain popular" tools only |
| Copyleaks | Yes, listed as a feature | None specific to paraphrasing | No paraphrase-specific evidence |
| Originality.ai | Yes, names QuillBot as an example | None specific on its product page | Built and tested for web content |
For how these tools score on ordinary, unparaphrased essays, our 12-tool AI detector comparison has our own test results.
What the research says about paraphrasing and AI detection
The best-known study is Krishna et al. (2023), "Paraphrasing evades detectors of AI-generated text". The authors built an 11-billion-parameter paraphraser called DIPPER and ran it over text from three language models. Paraphrasing evaded several detectors, including watermarking, GPTZero, DetectGPT and OpenAI's classifier.
Krishna et al. (2023) found that DIPPER paraphrasing dropped DetectGPT's detection accuracy from 70.3% to 4.6% at a 1% false positive rate.
The same paper offered a fix that only AI companies can run: compare a suspect text against a database of everything the model has generated. With 15 million stored generations, that retrieval defence caught 80% to 97% of paraphrased text while flagging 1% of human writing. No classroom detector works that way.
Sadasivan et al. (2023), "Can AI-Generated Text be Reliably Detected?", went further. Their recursive paraphrasing attack, run on passages of about 300 tokens, significantly reduced detection rates across watermarking, neural network, zero-shot and retrieval-based detectors, often with only slight loss of text quality. The authors also argue that detection gets harder as AI writing grows closer to human writing.
Both studies are from 2023 and tested research detectors and early commercial tools. Vendors have retrained since, which is why the vendor claims below exist. What has not appeared is an independent, peer-reviewed test of a commercial detector on paraphrased student essays. Until it does, treat paraphrase detection as unproven in classrooms.
How detectors handle paraphrased AI text
Turnitin
Turnitin's AI writing detection FAQ, updated October 7, 2026, says the detector can identify AI-generated text "modified by AI paraphraser or bypasser (also called humanizers) tools". The check runs automatically for institutions with the AI writing indicator enabled. Paraphrased passages are highlighted as part of the overall AI percentage, which counts text that was "likely AI-generated" and text that was also "likely AI-paraphrased".
Three limits matter for teachers. Paraphrase and bypasser detection works on English submissions only. Turnitin will not say which paraphrasing tools it was trained on, so "does it catch QuillBot?" has no official answer. And the indicator ignores papers under 300 words and hides scores between 1% and 19%. Our article on whether Turnitin detects ChatGPT covers the rest of its record.
GPTZero
GPTZero's support article says AI paraphrasing detection is in beta. When it finds paraphrased AI text with medium or high confidence, the scan shows "possible AI paraphrasing detected". In a November 2024 announcement, GPTZero reported about 95% accuracy on certain popular bypassing tools and said it mislabels human text as paraphrased in fewer than 1 in 1,000 cases.
GPTZero is careful about scope. It targets humanizers built to beat detectors, not writing aids such as Grammarly or plagiarism-avoidance paraphrasing. It also says its results "should not be used to academically punish". Our GPTZero review covers how it did on ordinary classroom essays.
Copyleaks
Copyleaks' AI detector page lists "Plagiarism & Paraphrased AI Detection" as a feature and claims over 99% accuracy with a 0.03% false positive rate overall. It does not publish a figure for paraphrased text specifically. Our Copyleaks accuracy review compares its claims with our results.
Originality.ai
Originality.ai's AI checker page says the model is trained on "adversarial data": its team tests tools that change or edit text, naming QuillBot and Grammarly, and retrains when one gets through. It describes itself as accurate on AI-paraphrased content and offers an "AI Allowance" setting for teachers who permit some AI paraphrasing. Our Originality.ai review notes that its testing has centred on web content rather than student writing.
Paraphraser vs humanizer: what is the difference?
A paraphraser rewrites text: simpler words, a different sentence order, a new tone. Students use paraphrasers on their own drafts, on sources, and on AI output. A humanizer is a paraphraser tuned for one job, making AI-written text score as human on detectors.
The line is blurry because many companies sell both, sometimes beside a detector. Grammarly lists a paraphrasing tool, an AI humanizer and an AI detector. So does Scribbr. Walter Writes offers an AI humanizer alongside an AI detector you can try free. For a teacher, the label on the tool matters less than what the student did with it.
| What the student did | Which check applies | How much a score tells you |
|---|---|---|
| Reworded a source without citing it | Plagiarism checker, plus your knowledge of the source | Moderate; heavy rewording lowers similarity |
| Pasted ChatGPT output unedited | AI detector | Fairly strong, but still not proof |
| Ran ChatGPT output through a paraphraser or humanizer | AI detector with paraphrase detection | Weak; vendor claims, little independent evidence |
| Polished their own draft with a paraphraser | Your AI policy; process evidence | Risky; may flag honest work |
Is paraphrasing with AI plagiarism?
It depends on what was paraphrased. If a student reworded someone else's article without citing it, that is plagiarism no matter which tool did the rewording. If a student submitted AI-generated text as their own work, most current school AI policies treat that as misconduct whether or not it was paraphrased afterwards. If a student used a paraphraser to tidy their own sentences, many policies allow it, provided it is disclosed.
Plagiarism checkers and AI detectors look for different evidence, which our explainer on plagiarism vs AI detection sets out. A paraphrased source can slip past a plagiarism checker while an AI detector has nothing to say about it, and the reverse. Our comparison of the best plagiarism and AI detection combo tools covers products that run both checks.
What a teacher can actually check
Because a paraphraser is designed to change exactly the signals detectors read, process evidence is more reliable than any score. These checks work regardless of which tool a student used.
- Version history. Google Docs and Microsoft Word keep revision history. A long essay that appears in one or two large pastes looks different from one built over several sessions.
- Drafts and notes. Ask for outlines, notes and earlier drafts as part of the assignment, not only after a flag.
- Writing-process tools. Grammarly Authorship sorts text by where it came from (typed, pasted or AI-generated), and GPTZero offers authorship and writing replay tools. These record the process rather than guess from the result.
- A short conversation. Ask the student to explain the argument, a source or a word choice. Our guide to oral defense assessment turns this into a routine rather than an accusation.
- Source checks. Look up two or three cited sources. Paraphrased AI text sometimes cites sources that do not say what the essay claims.
If a detector has already flagged a student who denies using AI, our guide on what to do when a student denies AI use walks through the next steps. For the wider picture of how students evade detection, see the bypass problem.
Frequently Asked Questions
Can Turnitin detect QuillBot?▼
Possibly, on English papers. Turnitin says its detector is trained on leading paraphraser and bypasser tools but will not name them, so it neither confirms nor denies QuillBot. Paraphrase detection is English only, and Turnitin never presents a score as proof.
Can GPTZero detect paraphrasing?▼
Sometimes. GPTZero says its paraphrase detection is in beta, catches text from certain popular bypassing tools with about 95% accuracy, and mislabels human text as paraphrased in under 0.1% of cases. Results show as "possible AI paraphrasing detected".
Can Copyleaks detect paraphrased AI text?▼
Copyleaks says yes: its AI detector page lists "Plagiarism & Paraphrased AI Detection". It has not published a separate accuracy figure for paraphrased text, so treat the general 99% claim as untested for this case.
Can Originality.ai detect QuillBot?▼
Originality.ai says it is accurate on AI-paraphrased content and names QuillBot as an example. Its team trains on adversarial data from tools that rewrite text. The figures come from the vendor, not from an independent test of student writing.
Is paraphrasing with AI plagiarism?▼
It depends on what was paraphrased and on your policy. Rewording someone else’s source without citing it is plagiarism whatever tool did the rewording. Submitting AI-written text as your own breaks most current AI policies even after paraphrasing.
Do plagiarism checkers catch paraphrased sources?▼
Partly. Checkers such as Originality.ai and Copyleaks say they detect paraphrased or translated plagiarism, but heavy rewording lowers similarity scores. A teacher who knows the source can often spot it faster than software.
What is the difference between a paraphraser and an AI humanizer?▼
A paraphraser rewrites text for clarity or style. A humanizer is built to make AI text pass AI detectors. GPTZero draws the same line: it targets bypassers and humanizers, not writing aids like Grammarly.
Can teachers tell if a student paraphrased ChatGPT?▼
Often, but not from a detector score alone. Version history, drafts, notes and a short conversation about the argument tell a teacher more than a percentage does. Our guide can teachers tell if you used ChatGPT covers the signs.
Does paraphrase detection work in other languages?▼
Mostly not yet. Turnitin states that its AI paraphrasing and bypasser detection is available for English submissions only, even though its AI writing detection covers seven languages.
The Bottom Line
Can paraphrasing tools be detected? Sometimes. Turnitin, GPTZero, Copyleaks and Originality.ai all say they now flag paraphrased AI text, but only GPTZero publishes a paraphrase-specific accuracy figure, and the independent research still shows paraphrasing can defeat detectors. A clean score on a paraphrased essay proves little, and a flag proves little more.
Design assignments so the writing process is visible, ask for drafts, and use detector results only as a reason to look closer.
