AI Detector FAQs: Turnitin, GPTZero and the Questions Students and Teachers Ask Most

Most people meet AI detectors through a single stressful moment: a score on their own work, a policy in a syllabus, a rumour in a group chat. The questions that follow are remarkably consistent — and remarkably badly served by folklore. This FAQ answers the dozen questions students, teachers and applicants actually ask, from the primary sources: Turnitin’s official FAQ, GPTZero’s explainer, the peer-reviewed studies, and Vanderbilt’s public policy reasoning. One boundary up front: several popular questions ask, in effect, how to disguise AI-written work so a checker misses it. That is academic dishonesty with extra steps, and we do not answer it. The legitimate questions — accuracy, fairness, what scores mean, what to do if you are flagged — are answered fully below, and the deeper evidence lives in the pillar, AI Detectors Explained.

What does an AI detector score actually mean?

On Turnitin, the figure is a percentage of sentences. Its FAQ explains that each sentence is scored between 0 and 1 in the context of the document (0 meaning very unlikely to be AI-written, 1 meaning very likely), and the headline score is the percentage of sentences that cross the AI threshold. GPTZero classifies at sentence and document level using a machine-learning model, reporting probabilities per sentence. In neither case does the number describe a measurement of your writing process — detectors analyse finished text for statistical resemblance to machine output, using signals like perplexity and burstiness explained in our mechanism guide. A 40% score means “40% of this text statistically resembles the machine writing in our models,” full stop. Everything beyond that sentence is interpretation, and interpretation is where institutions succeed or fail their students.

How accurate are Turnitin and GPTZero, really?

By their own published figures: Turnitin states a false positive rate below 1% for documents scoring over 20% AI, validated before each model release on more than 700,000 academic papers written before ChatGPT — and immediately adds that the score “should not be used as the sole basis for adverse actions” and that the tool “does not determine misconduct.” GPTZero claims 99% accuracy separating AI from human text with a 1% false positive rate, 95.7% AI-text identification on the adversarial RAID benchmark, and states in the same document that “no detector is perfect” and cannot provide definitive proof of authorship. Independent evidence is less flattering: a 2023 peer-reviewed test of 14 tools (Weber-Wulff et al.) found none above 80% accuracy on realistic material and worse performance on paraphrased text, while OpenAI withdrew its own classifier in 2023 over inaccuracy (it caught 26% of AI text and falsely labelled 9% of human text on its own test set). The reconciliation: detectors are strong on unedited chatbot output, weak on edited output, and measurably risky on unusual human writing. The full claim-by-claim audit is in how accurate are AI detectors?

Why was my own writing flagged?

The known, documented causes, in rough order of evidential support: predictable style — detectors key on low perplexity, and formulaic, template-following or heavily edited prose is predictable (Liang et al. found flagged essays had significantly lower perplexity); writing in an additional language — the Patterns study measured a 61.22% average misclassification rate for non-native TOEFL essays, against about 5.19% for US eighth-grade essays, because a careful second-language vocabulary is statistically narrow; short or partial samples — both vendors warn short text is harder to judge, and Turnitin requires at least 300 words of evaluated sentences and suppresses scores under 20%; machine assistance short of generation — translation, grammar tools and heavy autocomplete nudge text toward machine statistics, a grey zone the scores cannot express; and genre — lab reports, legal-style writing and other deliberately impersonal forms sit near machine texture by design. Turnitin notes it excludes lists, headers, references, tables, formulas and code from analysis — if your flagged document is mostly prose around those elements, the evaluated base may be small. Our dedicated analysis is AI detector false positives.

Can a teacher rely on the score alone?

No — and this is the rare FAQ answer where the vendor, the research and the dissenting university all say the same word. Turnitin: the score “should not be used as the sole basis for adverse actions against a student” and determinations of misconduct are the educator’s. GPTZero: results should be taken “with a grain of salt” and cannot offer “definitive proof.” Vanderbilt, disabling the tool: a 1% error rate across 75,000 submissions meant roughly 750 papers could be wrongly labelled, so detection was retired rather than trusted. A score can legitimately open a conversation; an honest case needs corroboration — drafts, version history, the student’s ability to discuss the work — because those are evidence about authorship, which a resemblance percentage is not. Teachers setting policy will find the institutional landscape, including how employers handle the same question with fewer guardrails, in what schools and employers actually do.

What should I do if I am falsely flagged?

Stay calm, get specific, and produce process. Ask what the evidence is beyond the percentage — which passages, which policy, what process follows. Offer your drafting history: version history in Google Docs or Word, earlier drafts, notes, sources — the artefacts of writing as it happened; they are the strongest authorship evidence that exists and the thing a detector never saw. Offer to discuss or reproduce the work — talking through your argument, or writing a short passage live, demonstrates command that a plagiarised submission rarely survives. Put the published record on the table, in writing: the vendor’s “not the sole basis” instruction, the Patterns bias findings if you write in an additional language, and Vanderbilt’s calculation are all citable, reputable sources — the Sources section below collects them. Use the formal appeal channel if the first conversation fails, and keep records of everything. Note the asymmetry to avoid: this advice is about establishing authorship of your own honest work. If AI tools were used in ways your institution prohibits, the honest route is disclosure under whatever amnesty or mitigation the policy provides — not score-shopping across detectors hoping for a lower number, which helps no one and convinces no one.

Do detectors store or train on my document?

Turnitin addresses this directly in its FAQ: documents submitted for the AI score are not stored and are not added to its content comparison databases, and after an instructor views the submission it cannot be re-processed through the AI check. For other tools, the answer lives in each product’s privacy terms — worth reading before pasting sensitive or unpublished work into a free web checker, since “free” detectors vary widely in what they retain. As a habit: institutional submissions through your school’s own system are governed by its vendor contracts; personal use of random checker sites is where caution belongs.

Are detectors biased against non-native English writers?

On the published evidence, yes — measurably, and by a mechanism experts can explain. Liang et al. (Patterns, 2023) found the seven detectors they tested misclassified non-native TOEFL essays at 61.22% on average while handling US school essays almost perfectly, traced the disparity to lower perplexity in second-language academic writing, and titled their conclusion plainly: the detectors are biased against non-native English writers. Turnitin says its training data includes second-language learner writing specifically to minimise bias, and the disparity is the central fairness problem the field acknowledges. Any institution running detection over a multilingual student body owns this finding whether it has read it or not — which is why fair-process design, covered in our false positives article, is not optional garnish on a detection policy.

What this means

If this FAQ has a single lesson, it is that the folklore around detectors — they know, they are always right, a flag is a verdict — is contradicted by every primary source in the field, including the companies selling the tools. The accurate mental model is duller and safer: detectors estimate stylistic resemblance, well enough to catch lazy copy-paste, badly enough that the creator of ChatGPT abandoned its own attempt, and unevenly enough that the errors land disproportionately on non-native writers and anyone trained into a predictable style. For US readers the practical posture follows. Students: write your own work and keep your drafts, because process is the only proof of authorship anyone can actually inspect. Teachers: the vendors have handed you the policy language — “not the sole basis” — and the research has handed you the error rates; a process built on both is defensible, and one built on neither is an accident in progress. Applicants and employees: know that workplace screening is the least regulated use of these tools, and that the same evidence base applies to it, whether or not your employer’s policy mentions it. The questions in this FAQ will keep being asked as long as percentages are attached to people’s writing. The answers will keep being the ones the sources give here: a score is a signal, fairness is a process, and authorship is proved the old way — by showing your work.

Parents and school boards, who increasingly field these questions at kitchen tables and public meetings, have a role in that settlement too. The questions worth asking any institution are the ones this FAQ is built from: What false positive rate does your vendor state, and over what population was it measured? What evidence, beyond the score, does your process require before a finding? What appeal exists, and has it ever reversed one? An institution that can answer those three questions has done the reading; one that cannot is running the Vanderbilt arithmetic blind, on other people’s children. The technology will keep changing — detectors will retrain, generators will adapt, percentages will fluctuate — but the governance questions are stable, and they were all, notably, answered in advance by the vendors’ own fine print and one university’s honest arithmetic. The sources below are short. Read them before you need them.

Bottom line

Every primary source in this field, including the sellers, contradicts the folklore. Detectors estimate style; a flag is the start of a fair process, not the end of one.

Read Next

• AI Detectors Explained: How AI Text Checkers Work and Why They Sometimes Get It Wrong
• How Accurate Are AI Detectors? What Studies and the Vendors’ Own Tests Actually Show

Sources

• Turnitin’s AI writing detection capabilities — FAQs (accessed 7 November 2026)
• How AI Detectors Work — GPTZero (accessed 7 November 2026)
• Liang et al. (2023), “GPT detectors are biased against non-native English writers,” Patterns 4(7) (accessed 7 November 2026)
• Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector — Vanderbilt University (accessed 7 November 2026)