Can Turnitin Detect QuillBot? AI Writing Tools, Turnitin and the IB Rules That Actually Matter

Picture two students in the same DP2 English class. Neither is a villain. Both are about to have a very bad fortnight.

The first student cheats and thinks she is safe.She generates a draft with a chatbot, runs it through a paraphraser on “Humanize” mode, reads it once, and submits. The output is clean, the vocabulary is impressive, and no plagiarism checker will match it to any source on the internet, because those exact words have never existed before. She believes she has won. What she has actually done is leave two separate signals — the statistical fingerprint of machine text, and the second fingerprint of a spinner applied on top of it — in a document her teacher has been reading drafts of for six months.

The second student cheats at nothing and gets accused anyway.English is his third language. He writes carefully and plainly, in short declarative sentences, with the vocabulary he is confident about. That style — low variability, predictable word choice — is precisely what a lot of detectors read as machine-generated. He submits an honest 900-word IA and gets an email asking him to come and discuss it.

Those two students need opposite advice, and almost every article on this subject only serves one of them. The first needs to understand what the systems can see. The second needs to understand that a number on a screen is not proof, and that he can prove authorship if he prepares to. This guide does both.

Start with the thing most students get backwards. The IB is not trying to ban these tools. Its published position, first issued in March 2023 and still maintained on ibo.org, opens like this:

The IB even puts these tools in the same category as the ones your school already hands you: it expects AI “will become part of our everyday lives—like spell checkers, translation software and calculators.” That comparison is genuinely useful, because it tells you where the line falls. A calculator is allowed. A calculator that sits the exam for you is not.

Then comes the sentence that decides almost every real case:

Read “even only in part” again, because that is the phrase that catches the paraphraser. If a tool rewrote your sentence, the sentence you submit was produced in part by that tool. The IB’s own guidance on AI, published alongside its academic integrity policy, is equally direct: students “should also be taught that AI-generated work included in a piece of assessment must be credited in the body of the text and appropriately referenced in the bibliography.”

There is one more line worth knowing, because it removes the get-out-of-jail move students reach for — the idea that if you cite the AI, you can hand in whatever it wrote. The IB closes that door in the same breath: “an essay which is predominantly quotes will not get many, if any, marks with an IB mark scheme.” Citing the machine keeps you honest. It does not get you marks.

Why the IB is not panicking

Dr Matt Glanville, the IB’s Director of Assessment, framed the organisation’s thinking in a piece for the IB community that is worth reading in full. His argument is that this is not a new problem wearing a new hat:

“This is not a new assessment problem… the problem of students getting someone else to write their work for them has existed for many years. Teachers and the IB regularly deal with cases of essays bought from the internet (from so called ‘essay mills’), completed by external tutors or even by family members.”

That is a more revealing quote than it looks. It tells you that the IB’s defence was never primarily technological. It was always process— and process is the thing that catches AI too. More on that next.

Students imagine a single scary machine that returns a verdict. What actually exists is four overlapping layers, and the software is the weakest of them.

Layer 1: The similarity report

The oldest layer, and the one most IB coursework passes through on its way to your teacher. It matches your text against web pages, publications and a vast archive of previously submitted student work. It does not care about AI at all. It catches copied sources, recycled coursework, and the essay your cousin submitted in 2023.

This is where students get a false sense of security about paraphrasers. A spun sentence will not match a source, so similarity stays low. But similarity was never the layer that was going to catch you.

Layer 2: The AI writing indicator

Turnitin’s AI writing detection returns a percentage estimate of how much of a document was likely machine-generated. Three practical details matter to you:

• It needs length. Turnitin requires at least 300 words of long-form prose, and it raised that floor from 150 words precisely because short pieces produced too many false alarms.
• You cannot see it. The AI indicator and report are shown to instructors, not students. You will not know your score unless your teacher shows you.
• It is not a verdict. Turnitin says so itself, which we will come back to.

Layer 3: AI paraphrasing detection — the QuillBot layer

This is the layer that answers the question in the title, and it is newer than most students realise. Turnitin added AI paraphrasing detection on top of its AI writing indicator, aimed at exactly one behaviour. In Turnitin’s own description of the problem:

“Students and researchers may use AI paraphrasing to modify AI-generated content, such as that produced by language models like ChatGPT, in an attempt to evade detection by AI detection software.”

The feature highlights AI-generated text that “may have been further modified using an AI paraphrasing tool”, in a separate colour, with its own percentage in the report. So the honest answer to can Turnitin detect QuillBot?is this: Turnitin does not have a QuillBot logo detector, and no vendor can promise to identify which product touched your text. What it has is a model trained to spot the specific signature of machine text that has been re-spun — and re-spinning machine text is the entire use case people reach for QuillBot for. The workflow is the thing being detected, not the brand.

Note the important limit, because it cuts the other way too: this layer looks for AI writing that was then paraphrased. It is not designed to flag your own honest prose because you once used a synonym tool. That is a real distinction, and it is why the next layer matters more than any of this.

Layer 4: The one that actually gets people — your teacher

Every piece of IB coursework is authenticated by a human being who has read your work all year. Glanville set out what that teacher is expected to be able to say, and it is a startlingly low-tech list:

• They have seen the work develop over time — “IB coursework is not designed to be completed in a single evening.”
• You can explain your work well enough to give confidence you created it.
• You are clear about which ideas are quoted and which are yours.
• The quality of the final piece matches what they would expect you to produce.

That last one is the trap door. A student who has written competent, slightly uneven 5-level prose all year and submits a flawless, oddly generic 7-level essay has not beaten a detector. They have failed a comparison a human made in four seconds. And the check-in meetings built into IA and Extended Essay supervision exist precisely so that this comparison happens repeatedly, on the record, before submission.

If you take one thing from this section: you are not trying to beat software. You are trying to be a person whose work is recognisably theirs. Those are completely different projects, and only one of them is achievable.

Below is the map. It is organised by what you do, not by which app you opened, because that is how the rules are written. The same product can sit in all three bands depending on which button you press — which is exactly why “is Grammarly allowed?” has no single answer.

Now the other student. The one who did nothing wrong.

Most AI detectors work, broadly, by measuring how predictable your writing is. Machine text tends to choose the unsurprising next word. So does the writing of someone composing carefully in a second or third language, using vocabulary they are sure of. The overlap is not a bug in one product; it is a structural property of the whole approach.

The landmark study here is Liang et al. (2023), published in Patterns by a Stanford team. They ran seven widely used GPT detectors over 91 human-written TOEFL essays and 88 essays by US eighth-graders. The detectors were near-perfect on the American students. On the TOEFL essays, the average false-positive rate was 61.22%. Nearly all of them — 89 of 91 essays, or 97.8% — were flagged as AI by at least one detector, and 18 essays were flagged by all seven.

Then they did something that should end the argument on its own. They asked a model to enrich the vocabulary of those same human-written essays, and the average false-positive rate fell from 61.22% to 11.77%. The detectors were not measuring authorship. They were measuring how fancy the word choice was.

Now consider who sits IB exams. The Diploma is a global programme, taught in international and bilingual schools, in which enormous numbers of students are writing assessed work in a language that is not their first. This is not a marginal risk for our community. It is a structural one.

The fair counter-argument, stated properly

Honesty requires the other side, and it is stronger than the alarmed version of this story usually admits. Turnitin’s detector was not among the seven tested by Liang et al., and Turnitin published its own evaluation in response, noting that the TOEFL essays in that study were all under 150 words — below the length at which Turnitin will make a prediction at all.

In its own testing, Turnitin says it validated against 800,000 academic writing samples written before ChatGPT existed, and then specifically compared second-language and first-language English writers. Its finding: for documents meeting the 300-word minimum, the difference in false-positive rate between the two groups was “small, and not statistically significant”. For documents that were too short, the gap was significant — which is why the minimum exists.

So the accurate summary is narrower and more useful than either camp’s headline: detector bias against second-language writers is well evidenced across the category, contested for Turnitin specifically, and worst on short pieces. If your school runs a detector over a 400-word reflection or a short commentary, you are in the least reliable part of the range.

What the vendor itself says its number means

Turnitin has been unusually candid about the limits of its own product, and these are the two sentences every IB student and parent should have to hand.

Turnitin also notes that its document-level guarantee applies to documents with 20% or more AI writing — below that threshold, it has said, false positives are more common. A score of “9% AI” on your IA is therefore close to meaningless, and a teacher treating it as evidence is misreading the product.

Some institutions have concluded the whole exercise is not worth it. Vanderbilt University disabled Turnitin’s AI detector in August 2023, doing the arithmetic publicly: it submitted 75,000 papers in 2022, so even a 1% false-positive rate implies roughly 750 students wrongly implicated in a single year. Whatever you think of the decision, the arithmetic is not in dispute.

Here is the uncomfortable asymmetry: a detector score takes a second to produce and can take weeks to argue with. You cannot control the score. You can control whether you have evidence.

Everything below costs you almost nothing and works whether you are protecting yourself from a false accusation or simply doing the work properly in the first place. Start it at the beginning of a piece, not when the email arrives.

First, know what is actually at stake, because the rumours are worse than the rules and the rules are bad enough.

The IB’s General regulations: Diploma Programme define academic misconduct as behaviour “(whether deliberate or inadvertent) that results in, or may result in, the candidate or any other candidate gaining an unfair advantage” — note inadvertent; intent is not a defence — and define plagiarism as “the representation, intentionally or unintentionally, of the ideas, words or work of another person without proper, clear and explicit acknowledgment”.

On consequences, the regulations distinguish between a lesser academic infringement, where no marks are awarded for the part of the component concerned but you remain eligible for a grade in the subject, and established misconduct, where no grade may be awarded in the subject at all. And because the Diploma requires a grade in every contributing subject, a single missing grade means no diploma. That is the real downside: not a lower score, but a hole where a subject used to be.

If a conversation is coming, this is the order to do things in.

Suppose none of the above applied. Suppose the detectors were switched off, your teacher never asked, and the risk was genuinely zero. There would still be a reason not to do it, and it is the reason that ought to matter most in your final two years of school.

A team at the MIT Media Lab ran a study in which participants wrote essays under three conditions — unaided, with a search engine, and with a language model — while their brain activity was recorded, across four months. The pattern they report is stark: “Brain-only participants exhibited the strongest, most distributed networks; Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity.” The model group also, memorably, “struggled to accurately quote their own work” and reported the lowest sense of ownership over it. The authors call the effect “cognitive debt”. It is a preprint rather than a peer-reviewed paper, with a modest sample, and it deserves to be read as suggestive rather than settled — but it names something every teacher recognises.

You do not have to accept the neuroscience to accept the mechanism, because the IB has already built the consequence into the assessment. Your IA is defended in supervision meetings. Your TOK essay grows out of an interactive oral. Your Extended Essay ends in a reflective conversation with your supervisor. Written exams — still handwritten or on-screen in a locked room, with no tools — are worth the majority of your points in most subjects.

The Diploma is, structurally, a two-year rehearsal for one week in May where nothing can help you but what is in your head. An outsourced IA in November is a loan against that week, and it is due in full.

Meanwhile the students who use these tools well are doing something quietly different. They are generating practice questions, arguing with a model about their own thesis, asking for the counterargument they did not want to hear. That is the same instinct that makes AI such rich material for a TOK essay: the interesting question was never “will I get caught”, it is “what is knowledge worth when the answer is free?”

Almost nobody reaches for a paraphraser at 11pm because they want to cheat. They reach for it because the paragraph is not working, they cannot see why, and there is nobody awake to ask. The tool is not solving a moral problem. It is solving a loneliness problem.

Which is why the durable answer is a person. Someone who reads the draft, tells you the argument collapses in paragraph three, shows you why, and makes you rewrite it in your own words — so that the version you submit is yours, defensible in a supervision meeting, and better than anything a model would have produced anyway.

That is the whole of what we do. Bespoke Learning matches students with IB-trained educators who know the coursework criteria, the academic honesty rules, and the difference between help that builds a writer and help that replaces one.