How Turnitin's AI writing detection works, and what to do if it flags you
By Vitalik Hakim · Published 20 September 2026 · Updated 18 October 2026 · Facts checked 18 October 2026
Turnitin's AI writing detection scores each sentence of the long-form prose in a submission and reports the share of sentences above its threshold as a percentage that only your instructor sees, unless they choose to share the report. Turnitin says that score must not be the sole basis for action against a student, so a flag should lead to a conversation about how you wrote the work, and the best thing you can bring to it is your notes, drafts and version history.
We don't run Turnitin, so we publish no pass rate for it. The detector we do measure is here, with dates and raw data.
What Turnitin's AI indicator is
Turnitin switched on AI writing detection on 4 April 2023. It is a separate measurement from the similarity report most students know: the similarity score matches your text against sources, while the AI percentage is "different from and independent of the similarity score". Since the free preview ended in December 2023, an institution only gets the AI indicator if it licenses Turnitin Originality on top of its usual product.
Students do not see it by default. In Turnitin's words, "The AI writing detection indicator and report are not visible to students. However, with the PDF download feature, instructors can download and share the AI report with students." Administrators can also switch the feature off, and a growing number of universities have done exactly that (see below).
Some institutions also use Turnitin Clarity, launched in July 2025: a writing space where students draft inside Turnitin, with an optional AI assistant, while it records how the text was written. It matters for the section on being flagged, because that record can work in your favour.
How it scores a submission
Turnitin describes the method plainly: a model gives every sentence a score between 0 and 1, "based on the sentence in context with the surrounding text", and "the document level AIW score is the percentage of sentences that are above the threshold." The model is a transformer trained on Turnitin's own collection of academic writing. Turnitin says it is "not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity'", though it learns similar patterns; how AI detectors work explains those ideas. On 14 October 2026 Turnitin updated the model for English to improve accuracy where human and AI text meet, and older reports need resubmitting to use it.
Only qualifying text counts: "prose sentences contained in a long-form writing format". Turnitin strips out lists, headers, references, data tables and anything that looks like formulas or code, and says the model does not reliably assess poetry, scripts, code, bullet points, tables or annotated bibliographies. Stripped text is left out of the percentage, so a slide outline or a bibliography is not scored.
- Length: from 300 to 30,000 words of prose. Shorter submissions get no AI score.
- Files: under 100 MB, as .docx, .pdf, .txt or .rtf.
- Languages: English, Spanish, Japanese, Arabic, and since 6 October 2026 Turkish, German and Dutch. Detection of AI-paraphrased and humanized text is English only.
How to read the report
The indicator shows one of four things. A blue percentage is a score. An asterisk ("*%") means the score was between 1% and 19%: since July 2024 Turnitin shows no number and no highlights in that range, because, as its own guide says, "there is a higher incidence of false positives when the percentage is between 0 and 19." Two grey dashes mean the file was not processed, and an exclamation mark means an error.
The categories have changed several times. In December 2023 Turnitin began detecting AI text run through "an AI word spinner"; in July 2024 the report split into "AI-generated only" and "AI-generated text that was AI-paraphrased", shown in purple. On 4 August 2026, for English, Turnitin went back to "a single, unified set of blue highlights", partly to "reduce the risk of unwarranted academic misconduct reviews". A guide or a screenshot that shows purple highlights is describing the old report.
Read the number as an estimate with a margin. Turnitin's own example: "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing." The margin runs both ways, which is why it tells instructors to treat the score as one data point.
Turnitin and humanizer tools
Since 27 August 2025 Turnitin has run a feature it calls "AI bypasser detection", "designed to help educators detect text that may have been intentionally modified by AI humanizer tools". Text it attributes to such tools is counted in the "AI-generated only" figure. Turnitin will not say which tools it looks for: "we're unable to disclose the names of these tools."
AntiGPT is a humanizer, so this is worth saying plainly. We do not run Turnitin, we publish no figures for it, and we cannot tell you how it will score any text, including text rewritten with AntiGPT. If your course does not allow AI, rewriting AI text with any tool is still using AI, and it is the very thing this feature was built to look for. Where your course does allow AI help, follow its rules on disclosure.
Grammar tools are a different case. Turnitin says that "in most cases" changes made by Grammarly and other grammar checkers were not flagged, but that this "excludes content generated by Grammarly's generative AI-powered features, including draft generation, paraphrasing, summarizing", which "will likely be flagged".
False positives: what Turnitin says, and what others found
Turnitin's headline claim is a false-positive rate "under 1% for documents with over 20% of AI writing", which it explains as flagging one in every 100 fully human documents. It tests each model update on more than 700,000 academic papers written before ChatGPT existed. In 2023 it also put its sentence-level false-positive rate at about 4%, mostly in human sentences next to AI ones. It lists the kinds of human writing most likely to be flagged: text "without a lot of structural variation", text that repeats itself, and text "paraphrased without developing new ideas". That describes a lot of careful, formulaic academic writing.
One percent sounds small until you count. Vanderbilt University worked out that at 1%, its 75,000 papers a year would mean about 750 students wrongly flagged, and switched the detector off in August 2023. Others followed, including Montclair State (2023), the University of Waterloo (2025, after its own testing found "in more than one instance the product flagged human written text as 100% generated by AI"), Curtin University in Australia (from January 2026) and the University of the Free State in South Africa (from July 2026). Yale's teaching centre says the feature is "currently disabled" there.
Independent tests are mixed. A 2023 study of 14 detectors in the International Journal for Educational Integrity ranked Turnitin first, at 76% accuracy, with no false flags on its 18 human-written texts, while concluding that all the tools were "neither accurate nor reliable". A Washington Post test in 2023 found Turnitin wrong at least in part on over half of 16 samples, including an essay written entirely by a student. On non-native English writers, a 2023 Stanford study found seven other detectors heavily biased against them; Turnitin says its own research found no significant bias for texts over 300 words, but that research is not peer reviewed.
Our own benchmark cannot test Turnitin, and its human passages are under Turnitin's 300-word minimum, so Turnitin would not score them at all. It shows the general failure in a different detector: famous human writing, scored as likely AI, week after week.
What our own benchmark shows
Six passages written by people, long before AI writing tools existed, scored by the Hello-SimpleAI RoBERTa detector (open model) in our weekly benchmark. Run of 16 November 2026. "Likely AI" means a score of 60% or more.
| Passage | Words | Scored as AI | Verdict |
|---|---|---|---|
| Abraham Lincoln, the Gettysburg Address (1863) | 102 | 99% | Likely AI |
| Mark Twain, "Advice to Youth" (1882) | 70 | 87% | Likely AI |
| Frederick Douglass, Narrative of the Life of Frederick Douglass (1845) | 74 | 30% | Uncertain |
| Jane Austen, Pride and Prejudice (1813) | 70 | 2% | Likely human |
| Henry David Thoreau, Walden (1854) | 99 | 0% | Likely human |
| W. E. B. Du Bois, The Souls of Black Folk (1903) | 72 | 0% | Likely human |
If Turnitin flags your work
Turnitin's guidance to instructors is on your side here. It says the score "should not be used as the sole basis for adverse actions against a student" and that "Turnitin does not make a determination of misconduct". It suggests instructors open with questions like "Explain your process for completing this assignment" and, if the evidence is unclear after a review, "give the student the benefit of the doubt". The steps that help most:
- Ask to see the AI report itself, not just the number. Instructors can download it as a PDF and share it. Check whether the score is a real percentage or an asterisk, and which passages are highlighted.
- Bring your process. Turnitin itself tells students to "be prepared to share evidence: share any notes, outlines, version histories, or drafts". Google Docs keeps version history automatically, and Word does when the file is saved to OneDrive; open it before the meeting.
- If you wrote in Turnitin Clarity, ask for the Writing Report. It records typing, revisions and paste events, and playback of how the essay grew is strong evidence that you wrote it.
- Say what tools you used. A grammar checker is usually fine by Turnitin's own account; generative features, paraphrasers and translation tools are worth mentioning before anyone asks.
- Be ready to talk about the work: your argument, why you chose your sources, what you cut. Then use your institution's academic integrity process. Turnitin has no appeal route of its own; it says "the final decision on whether any misconduct has occurred rests with the reviewer/instructor."
If English is not your first language
Detectors have a documented record of flagging non-native writers more often, because careful second-language English tends to be more predictable. Turnitin disputes that this applies to its tool for long texts, but the safest course is the same either way: keep drafts, keep your notes in whatever language you think in, and say if you used translation. Our guide for non-native English writers covers the problem in more depth, and if you are accused of using AI covers the general steps.
Where AntiGPT fits, honestly
AntiGPT's detector is Hello-SimpleAI's open RoBERTa classifier, not Turnitin, and its score will not predict Turnitin's. What it can show you is how predictable your writing reads to one detector, before and after you edit it, as feedback rather than a forecast.
If your course allows AI assistance, AntiGPT helps the result read in your own voice; our guide to essays and coursework covers using it within the rules, citations and disclosure included. If your course does not allow AI, the honest advice is to write the essay yourself and keep the drafts. That is also your best defence if a detector gets it wrong.
Questions
Can students see their Turnitin AI score?
Not by default. The AI indicator and report are visible only to instructors and administrators, though instructors can download the report as a PDF and share it with you. You can ask to see it if your work is questioned.
What does an asterisk mean on Turnitin's AI score?
It means the score was between 1% and 19%. Since July 2024 Turnitin shows no number and no highlights in that range, because it says false positives are more common there.
Does Turnitin flag Grammarly?
Turnitin says grammar corrections from Grammarly and similar tools are usually not flagged, but text produced by Grammarly's generative features, such as drafting, paraphrasing or summarising, will likely be flagged.
Can Turnitin detect AI text that has been paraphrased or humanized?
Turnitin says it can. It added detection of AI-paraphrased text in 2023 and 2024, and since August 2025 it looks for text modified by humanizer tools, in English only, without naming the tools. We do not run Turnitin and cannot say how it scores any particular text.
How accurate is Turnitin's AI detector?
Turnitin reports a false-positive rate under 1% for documents it scores above 20% AI. Independent tests have been mixed, and several universities, including Vanderbilt, Waterloo and the University of the Free State, have switched it off over reliability concerns.
How does Turnitin's AI detection work?
It takes the long-form prose of a submission, at least 300 words and at most 30,000, after stripping out lists, tables, references and code. A model scores every sentence between 0 and 1 in the context of the text around it, and the report shows the percentage of sentences above Turnitin's threshold. Scores from 1% to 19% are shown only as an asterisk.
Sources
- Turnitin: AI writing detection FAQ (edited 14 October 2026)
- Turnitin: using the AI writing report (edited 26 August 2026)
- Turnitin: AI writing detection model release notes
- Turnitin: file requirements for AI writing detection
- Turnitin: unified AI writing highlights (August 2026)
- Turnitin press release: expanded capabilities for AI humanizer tools (August 2025)
- Turnitin: preparing for and discussing false positives
- Turnitin Clarity: what the writing report records
- Vanderbilt University: why we are disabling Turnitin's AI detector (August 2023)
- University of Waterloo: discontinuing Turnitin's AI detection
- University of the Free State: shift in academic integrity approach (May 2026)
- Weber-Wulff et al., Testing of detection tools for AI-generated text (2023)
- Liang et al., GPT detectors are biased against non-native English writers (2023)
- Turnitin: research on English language learners (October 2023)
Facts about Turnitin are taken from its own published pages, last checked 18 October 2026. Vendors change their products without notice; tell us at support@antigpt.app if something here is out of date.
Check your own text
Paste a paragraph you drafted with AI help. You will see how a detector scores it, then you can rewrite it in your own voice. Free: 300 words a week, sign in with Google; your text is kept while you do.
Detector scores are estimates from third-party classifiers and change as those tools update. No text can be guaranteed to pass any detector. Use this to improve your writing, and follow your institution's or employer's rules on AI use.
Related
- How GPTZero works, what its scores mean, and what to do if it flags your writingHow GPTZero classifies text as human, AI or mixed, what its percentage and confidence levels mean, the 2026 model, why it flags human writing, and what to do if it flags yours.
- How Originality.ai detects AI writing, what its score means, and what to do if it flags youWhat an Originality.ai score really means (a confidence level, not a share of AI text), its 2026 models, why it flags human writing, and what to do if a client flags your work.
- How the Copyleaks AI Detector works, what its results mean, and what to do if it flags youHow Copyleaks detects AI writing, what its percentage and highlighting mean, its sensitivity levels, why it flags human writing, and what to do if your school's Copyleaks check flags you.
All AI detector guides · how our own detector results are measured