Turnitin flagged your essay? What Turnitin itself says that score means

Read the manual your accuser may not have

If a Turnitin AI score has just landed on your essay, the most useful document in your defense is not a character reference or a detector-bashing blog post. It is Turnitin's own documentation — which is far more cautious about the score than most people who wield it.

This article does one thing: it walks through what Turnitin publicly says its AI writing score means, caveat by caveat, with the quotes you can cite. Every claim below comes from Turnitin's own blog, FAQ, or product guides, as of the dates in the sources list. Nothing here requires you to argue that the tool is bad — only that it be read the way its maker says to read it.

What the number actually claims

The score is a prediction: the model's estimate of what percentage of the qualifying prose in your document resembles AI-generated text. Two words in that sentence carry caveats most readers miss.

"Prediction." Turnitin's guide for instructors states it plainly: "Our AI writing detection model may not always be accurate… so it should not be used as the sole basis for adverse actions against a student." And the company's false-positives blog goes further on what the tool is for: "Turnitin does not make a determination of misconduct… rather, we provide data for educators to make an informed decision." The vendor positions the score as an input to human judgment — not a verdict. If it is being treated as a verdict, that use contradicts the manufacturer's instructions.

"Qualifying prose." The model does not score your whole document. By Turnitin's own description, it "eliminates text strings that it predicts are not actual sentences, such as lists, headers, references, data tables, and text that appears to be formulas or code." Separately, the company says the model "does not reliably detect AI-generated text in the form of non-prose, such as poetry, scripts, or code, nor does it detect short-form/unconventional writing such as bullet points, tables, or annotated bibliographies." A "98%" on a document that is half tables and citations is 98% of something much smaller than your essay.

The false-positive rate has a conditional clause

The headline accuracy claim is usually repeated as "less than 1% false positives." Here is the sentence as Turnitin's FAQ actually writes it:

"We strive to maximize the effectiveness of our detector while keeping our false positive rate — incorrectly identifying fully human-written text as AI-generated — under 1% for documents with over 20% of AI writing."

That bolded qualifier (emphasis ours) matters. The under-1% figure is scoped, by the vendor, to documents the model scores above a threshold — it is not an unconditional promise about every flag. Turnitin knows the low end is noisier, and says so: "In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing.'"

And even a true sub-1% rate is not small in a university. Vanderbilt — which disabled the detector — did the arithmetic in its public explanation: at its 2022 submission volume of around 75,000 papers, a 1% false-positive rate would have meant roughly 750 students' papers incorrectly flagged. "Sub-1%" and "hundreds of false accusations a year" are the same number wearing different clothes.

The asterisk: what Turnitin did to scores under 20%

In July 2024, Turnitin stopped showing numerical scores below 20% at all. From the product guide: "the AI indicator will display an asterisk (*) for percentages between 0 and 20 to call attention to the fact that the score is less reliable." The FAQ describes the same policy: "To avoid potential incidence of false positives, no score or highlights are attributed for AI detection scores in the 1% to 19% range."

Sit with what that means. The vendor looked at its own low-range scores and concluded they were too unreliable to display as numbers — replacing them with a symbol whose documented meaning is "less reliable." If your case involves a score under 20% (or a report generated before July 8, 2024, which may still show a low numerical score — Turnitin's guide notes this explicitly), the vendor's current position is that such numbers should not be surfaced as percentages at all.

The other eligibility rules

Buried in the file requirements are rules that decide whether the model should have scored your document in the first place:

• Minimum length. "Depending on the language, we do not process documents when the total number of words in evaluated sentences is under 300 words." The file requirements call for "at least 300 words of prose text in a long-form writing format" — and remember, lists, headers, and references don't count toward prose.
• Maximum length. Files "must not exceed 30,000 words."
• Language. Detection covers long-form writing in seven languages (English, Spanish, Japanese, Modern Arabic, Dutch, German, Turkish) — and Turnitin's "AI paraphrasing and bypassing detection capabilities are only available for English submissions."
• Training data. The model was tested against academic papers "written before the release of ChatGPT" — over 700,000 of them, per the FAQ. That is a reasonable benchmark, and also a reminder of what the tool is: a statistical comparison against a corpus, not an eyewitness.

If your flagged document is short, list-heavy, non-English, or creative non-prose, these are not technicalities — they are the vendor's own stated boundaries of the tool's competence.

Who lands in the error bar

False positives are not distributed evenly. A Stanford-led study (Liang et al., 2023) tested seven GPT detectors on essays by non-native English speakers and found their genuine writing falsely flagged 61.3% of the time on average — while the same detectors handled native speakers' essays nearly perfectly. The study covers GPT detectors as a class rather than Turnitin's current model specifically, but the mechanism it identifies — detectors penalizing the more uniform, lower-perplexity prose that careful second-language writers (and careful writers generally) produce — is the same mechanism behind every "your writing was too clean" flag. The full story of why honest work gets flagged is in the false accusation playbook.

How to use all of this

If you are responding to a flag right now:

1. Quote the vendor before you argue anything else. "Turnitin's own guidance states the score 'should not be used as the sole basis for adverse actions against a student'" is the strongest opening sentence available to you, because it reframes the question from your honesty to the process. The step-by-step playbook has a full email template built around exactly this move.
2. Check the eligibility rules against your document. Under 300 words of prose? Score under 20%, or an old report showing one? Heavy tables, citations, or bullet points? Non-English? Each is a vendor-documented reason the score deserves less weight — cite the specific rule.
3. Bring affirmative evidence, not just doubt. Undermining the score gets you to "unproven," not "cleared." Drafts, version history, notes, and your ability to talk through the work get you the rest of the way — here is every type of evidence, ranked.

And if this is the second time a detector has put you in this position: the only evidence that exists before an accusation is a record you keep while writing. That is what writing provenance is — and why a growing, verified list of universities has concluded that scores like this one shouldn't be the arbiter at all.

The one-paragraph version

Turnitin says its AI score is a prediction that "may not always be accurate," must not be "the sole basis" for action, and makes no "determination of misconduct." Its under-1% false-positive claim is scoped to documents scoring over 20% AI; below 20% it now shows an asterisk because those scores are "less reliable." It skips lists, tables, references, and code; requires 300 words of long-form prose; and covers seven languages. None of that is a critic's summary — it is the manufacturer's label, and any fair process has to read it.