Does Turnitin Detect QuillBot? Why Paraphrasers Fail AI Detection

Often, yes. When the text started as AI output, Turnitin now reports QuillBot-style paraphrasing in a category of its own, as described below. Paraphrasing tools in general fail against AI detection for a simple reason: they change words, while detectors respond to patterns that live above the word level. A synonym-swap tool can replace "significant" with "notable" in every sentence and leave the sentence length, clause order, rhythm and overall predictability almost untouched. When it does change enough to move a score, it often damages meaning or terminology along the way.

Detector vendors have also responded directly. Turnitin's documentation now describes a separate category for text that was likely AI-generated and then modified by an AI paraphrasing tool or word spinner, naming QuillBot as an example (Turnitin guides). This post explains what detectors actually react to, why spinners struggle, and how selective, meaning-preserving revision differs.

What a detector actually responds to

Detectors do not keep a list of "AI words". They score how predictable the text is to a language model, how evenly that predictability is spread across sentences, and, in trained classifiers, many subtler regularities learned from examples. We walk through the methods in how AI detectors work, but three signals matter most here.

• Predictability. Does each next token look like what a model would have chosen? Swapping one adjective for a near-synonym usually keeps the token about as likely as before.
• Structure. Topic sentence, supporting clause, "This suggests that" summary. Template paragraphs keep their skeleton after word-level edits.
• Rhythm. Sentence lengths and clause patterns that stay uniform across a paragraph. A spinner rarely splits or merges sentences.

So a spinner works on the one layer detectors weight least, and leaves the layers they weight most.

Why synonym swapping backfires

Classic spinners pick replacements from a thesaurus or a small model, with little sense of context. In academic text that creates two problems at once.

First, it can raise suspicion instead of lowering it. Odd word choices in an otherwise templated sentence are themselves a pattern, and trained classifiers can learn it. That is how a tool like Turnitin can report a paraphrased category at all.

Second, it breaks terminology. Research fields depend on fixed terms: "confidence interval", "random forest", "informed consent", "grounded theory". A thesaurus does not know they are fixed. Cabanac, Labbé and Magazinov documented this damage in published papers and named it "tortured phrases", giving "counterfeit consciousness" for "artificial intelligence" as an example (Cabanac et al., 2021). They found these phrases concentrated in the abstracts of one established journal, which shows how visible such substitutions are to an expert reader.

Before

We fitted a random forest model and report the mean absolute error on the held-out test set.

After

We fitted a haphazard woodland model and report the average total blunder on the kept-back test set.

The sentence structure is identical. The rhythm is identical. Only the terms changed, and every change is wrong. A reviewer would stop reading here.

When heavier paraphrasing does move scores

Stronger paraphrasers exist, and research shows they can lower detection rates a lot. Krishna and colleagues trained DIPPER, a paraphraser that can reorder content and vary wording with controllable intensity. Paraphrasing with it dropped DetectGPT's detection rate from 70.3% to 4.6% at a 1% false positive rate, and it also hurt watermark-based and commercial detectors (Krishna et al., NeurIPS 2023). Sadasivan et al. went further with recursive paraphrasing, applying a paraphraser several times in a row (Sadasivan et al., TMLR).

• Original model output70.3%
• After DIPPER paraphrase4.6%

Source: Krishna et al., "Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense", NeurIPS 2023

Two details from that literature matter for academic writers. The paraphrasers that worked changed structure and order, not just vocabulary, which supports the point above. And these were research models evaluated on general text, where the paraphrase only needed to stay close in meaning. A journal article has tighter constraints: a hedge, a unit or a citation that moves one clause over can change what the paper claims. Detectors also adapt. Krishna et al. proposed a retrieval defense in the same paper, and the RAID benchmark shows detector results swinging widely across attacks and settings (Dugan et al., ACL 2024). See our explainer on the DIPPER paper for more.

Does Turnitin detect QuillBot paraphrasing?

For AI-generated text, Turnitin says yes. Its AI Writing Report guide describes a separate highlight for text that was likely AI-generated and then modified by an AI paraphrasing tool or word spinner, naming QuillBot as an example. Since August 2025 Turnitin also looks for text passed through AI bypasser tools. Both checks run on English submissions only.

That was not always so. In the 2023 test by Weber-Wulff and colleagues, ChatGPT text run once through QuillBot on default settings was classified correctly only 26% of the time on average across 14 tools, Turnitin among them. Turnitin's paraphrase category came later, in a July 2024 update.

Two limits are worth knowing:

• Turnitin describes this category for text that was AI-generated first. It does not publish figures for human writing that was run through a paraphraser.
• A highlight is a prediction. Turnitin says its AI score should not be the sole basis for action against a student.

Humanizers raise the same question in a different form; see can Turnitin detect humanized text?

The real cost: meaning drift

Whole-document paraphrasing treats every sentence as a target. That includes sentences you wrote yourself, sentences no detector flagged, and sentences carrying exact claims. The more text a tool rewrites, the more chances it has to:

• turn "may be associated with" into "causes" (hedge drift),
• move a citation away from the claim it supports,
• round 0.047 to "about 0.05" or drop a unit,
• replace a defined term with a looser everyday word.

Each of these is easy to miss on a fast read and costly in peer review. Weber-Wulff and colleagues, testing 14 detection tools, found that content obfuscation, including manual edits and machine paraphrasing, significantly worsened tool performance (Weber-Wulff et al., 2023). That result is often quoted as a reason to paraphrase. Read it the other way: if detection is that unstable, the only thing you control for certain is whether your paper still says what you meant. We cover the review side in preserving meaning and citations when humanizing.

What selective, meaning-preserving rewriting does differently

A better approach starts from the detector's view of the draft rather than from the whole text. Instead of rewriting everything, it asks which sentences actually read as machine-like, and works only on those.

1. ScanScore each sentence and find the ones detectors flag.
2. LockMark numbers, units, citations, hedges and field terms as fixed.
3. Rewrite flagged sentences onlyChange structure and rhythm where needed, keeping the claim and its strength.
4. Re-checkScore the new version on the same measure as the original.
5. Author reviewRead each change next to the original before accepting it.

This differs from a spinner in three ways. It edits the layers detectors respond to (structure and rhythm) instead of vocabulary alone. It leaves human-reading sentences untouched, which keeps your voice and limits the area where errors can creep in. And it measures before and after, so you can see whether the change helped instead of guessing.

It is still revision, not magic. Some sentences will need a second pass, and the author remains responsible for every word. If your course or journal has rules on AI assistance, follow them; our guide on questions to ask before using AI on an assignment is a good starting point, and AI humanizer vs paraphraser vs grammar checker compares the tool types side by side.

How Academic Humanizer approaches it

Academic Humanizer follows the selective model. The Humanizer scans your draft for sentences that AI detectors flag, rewrites only those, and returns every other sentence exactly as you wrote it, with argument, numbers, citations and field terminology kept in place. If nothing is flagged, the output equals the input and no credits are charged. The AI Detection Report pairs each flagged sentence with its rewrite and compares both versions on the share of sentences read as human, checked against detectors such as Turnitin, GPTZero, Copyleaks and Originality.ai. You still read and approve the final text.