If you have ever worried that a detector might flag your own writing as AI-generated, you are not paranoid — you are reading the evidence correctly. AI detectors are real tools, universities really use them, and they really do make mistakes. The question is how often, for whom, and what a score actually means.
This guide separates what vendors claim from what independent research found, with the bias against non-native English writers front and centre — because for most students in the EU, that is the part that matters most.
How a detector actually works
A detector does not know whether AI wrote your essay. It measures statistical properties of the text — mainly perplexity (how predictable each word is) and burstiness (how much sentence length and structure varies). AI-generated text tends to be predictable and uniform: the model picks the most likely next word, and its sentences are a similar length. Human writing is messier: surprising word choices, short sentences next to long ones, an unusual turn of phrase.
That is the whole method, and it is also the whole problem. Anything that makes your writing more predictable — writing in a second language, writing formal academic prose, writing in a rigid template — makes it look more like AI output. And anything that makes AI output less predictable — asking it to vary sentence length, paraphrasing it, editing it by hand — makes it look more human. The detector is measuring a style, not an author. Once you understand that, the false positives stop being mysterious: they are what the method predicts.
What Turnitin claims
Turnitin is the detector that matters most, because it is the one wired into university workflows. Its claims, from its own documentation:
- Document-level false-positive rate: under 1% — but only for documents where 20% or more of the writing is AI-generated. That scope condition matters: the headline figure does not cover the papers most students submit.
- Sentence-level false-positive rate: about 4% — roughly one highlighted sentence in 25 may have been written by a person. Individual sentences are harder to classify than whole documents, and this is why passage highlights deserve more scepticism than the document figure.
- Scores between 1% and 19% are suppressed entirely and shown as an asterisk, because Turnitin's own internal testing found a higher false-positive risk in that band. Since July 2024, it attributes no score above 0% and below 20%.
- The score "should not be used as the sole basis for adverse action against a student." That is Turnitin's own published guidance, in its August 2026 update. Turnitin states that it does not make a determination of misconduct; it provides data for the educator to interpret.
Two more facts from Turnitin's own pages: the detector only works on prose sentences in long-form writing of at least 300 words — it does not evaluate code, equations or short answers — and it flags only at roughly 98% confidence, which by Turnitin's own estimate means it will miss up to 15% of AI-written text by design. The low false-positive rate is purchased with false negatives.
These are the vendor's measurements on the vendor's benchmarks. They are real, and they are also the best-case numbers.
What independent research found
The benchmark study is Weber-Wulff et al. (2023, International Journal for Educational Integrity), which tested 12 public tools plus two commercial systems, Turnitin and PlagiarismCheck. All fourteen tools scored below 80% real-world accuracy — below the reliability the authors considered necessary for misconduct decisions — and accuracy dropped sharply on paraphrased and manually edited AI text. A detector that catches raw ChatGPT output will often miss the same text after a student has rewritten it.
The most recent peer-reviewed work, from 2026, finds a pattern that should worry students and lecturers equally: false positives have largely disappeared for several tools, while false negatives have become severe. In one 2026 study (Van Vlasselaer et al., International Journal of Educational Integrity, June 2026), the tools found no false positives on fully human papers — but missed all 40 fully AI-generated papers, because their scores fell below the 20% threshold. Turnitin classified 100% of the fully AI-generated papers as false negatives. The detectors are getting better at not accusing the innocent, and worse at catching the guilty.
And the "98% accurate" figure you will see quoted everywhere? It comes from a 2023 vendor whitepaper, repeated by affiliates. Do not present it as a 2026 fact. It is not.
The bias against non-native English writers
This is the most important part of this guide, and the most solidly evidenced.
In 2023, a Stanford team (Liang, Yuksekgonul, Mao, Wu and Zou, published in Patterns, a Cell Press journal) tested seven widely used GPT detectors against 91 essays written by non-native English speakers for the TOEFL exam and 88 essays written by native-English-speaking US eighth-graders. The detectors falsely flagged 61.3% of the non-native essays as AI-generated, while scoring near-perfect accuracy on the native-speaker set. Eighteen of the 91 TOEFL essays — one in five — were unanimously misclassified by all seven detectors. And 89 of 91 were flagged by at least one detector.
The mechanism is structural, not incidental. Non-native English writers naturally use lower lexical diversity, simpler syntactic patterns and more predictable word choices — the same statistical properties detectors use as their primary signal for AI text. Detectors trained to spot AI patterns end up flagging anything that looks predictable, regardless of who wrote it. The Stanford authors put it plainly: they "strongly caution against the use of GPT detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers."
Has the gap closed since 2023? Partly, not fully. Independent re-testing of an updated GPTZero model in 2024 found a 7.7% false-positive rate on the same TOEFL benchmark — down substantially from the original cross-detector average, but not at parity. A February 2026 study (Al Ali, Libovický and Helcl) found a commercial detector with a 23.1% false-positive rate on non-native TOEFL essays against 0.0% on native essays. The bias has shrunk; it has not disappeared.
Turnitin disputes the bias for its own tool: it has published internal testing claiming a 0.014 false-positive rate for English Language Learner writing versus 0.013 for native-speaker writing — a gap its Chief Product Officer describes as not statistically significant. That is Turnitin's own measurement. No detector has published independently verified parity between native and non-native English writers.
One more finding that should change how you write: in a 2025 study, researchers took human-written research letters and used ChatGPT to polish them for readability and flow, with no factual changes. Before polishing, 97% to 100% were classified as human-written. After polishing, 75% to 85% were flagged as AI-generated, with 15% to 25% labelled at high confidence. Polishing your own writing with AI can make it look like AI wrote it.
What universities actually do with a score
A score is not a verdict. Turnitin's own guidance says to use the report to start a conversation, not to draw a conclusion — and universities are split on the tool. Vanderbilt University disabled Turnitin's AI detector in August 2023, citing reliability and false-positive concerns. Several UK universities have switched it off. Others run it routinely. And at Washington State, a third of AI cases built on detection alone ended in a finding of not responsible.
What matters in practice: students cannot see their own Turnitin AI score — only instructors and administrators can. A score below 20% shows as an asterisk, not a number. And a high score is a reason for your lecturer to look closer — to pull your drafts, check your edit history, and have a conversation — not proof of anything on its own.
What to do if you are flagged
If a detector flags your work and you wrote it yourself, do not panic — and do not stay silent. Three things protect you:
Keep your drafts. Every version of your essay, from the first outline to the final draft, is evidence that you wrote it. Save them with dates. Our guide on how universities detect AI writing explains why process evidence matters more than any score.
Keep your notes and sources. The reading notes, the highlighted PDFs, the half-formed outline — all of it shows the work behind the words. A student who can show their notes can show their thinking.
Be able to explain your work orally. If your lecturer asks you to walk through your argument, your sources and your choices, you should be able to do it. That is the strongest evidence there is, and it is the reason universities increasingly use oral defences alongside detection software.
And if you did use AI — for brainstorming, editing or drafting — say so. Disclose it the way your university asks, keep a record of your prompts, and make sure the ideas and the final wording are yours. Honest disclosure is always safer than a clean detector score built on silence.
Frequently asked questions
Are AI detectors accurate?
Not accurate enough to decide a misconduct case on their own. Turnitin claims under 1% document-level false positives (only for papers with 20% or more AI writing) and about 4% at sentence level. Independent tests found all 14 tools tested below 80% real-world accuracy, and 2026 studies show detectors now miss most fully AI-generated papers.
Can a detector flag my own writing as AI?
Yes — especially if English is not your first language. A Stanford study found detectors falsely flagged 61.3% of TOEFL essays by non-native English writers. The gap has narrowed since 2023 but has not closed. Polishing your own writing with AI can also trigger a flag.
What does a Turnitin AI score of 20% mean?
It means Turnitin estimates that 20% or more of the document was AI-generated. Scores below 20% are shown as an asterisk, not a number, because Turnitin found a higher false-positive risk in that range. The score is not a determination of misconduct.
Can I see my own Turnitin AI score?
No. Students cannot see their own score — only instructors and administrators can. If you are worried, ask your lecturer what the report showed and what evidence they want from you.
Do detectors work on paraphrased AI text?
Poorly. Accuracy drops sharply on paraphrased and manually edited AI text, and simple modifications can cut detection rates by more than half. That is why universities increasingly rely on drafts, edit history and oral defences rather than detection alone.
The bottom line
AI detectors are real, and they are not accurate enough to trust as the sole basis for a misconduct decision — Turnitin itself says so. Their headline figures are measured under conditions that do not resemble a messy classroom submission mix, and they carry a documented bias against non-native English writers that has narrowed but not closed. If you write in English as a second language, the risk is not theoretical: it is the best-documented finding in the field. Keep your drafts, keep your notes, disclose your AI use honestly, and be able to explain your work. That is the evidence that actually protects you — not a clean detector score.
Related guides
- How Universities Detect AI Writing in 2026 — and How to Stay Honest
- How to Cite AI Tools in APA, MLA and Chicago — Plus the Disclosure Statement EU Universities Want
- Is Using AI Cheating? What EU Universities Actually Say
- How to Write an Essay with AI: A Step-by-Step Guide for Students
Accuracy figures and study findings checked against Turnitin's own documentation and the cited peer-reviewed studies (Liang et al., Patterns 2023; Weber-Wulff et al., 2023; Van Vlasselaer et al., Int. J. Educ. Integrity 2026; Al Ali et al., 2026) in November 2026. Detector performance changes with each model update — treat any accuracy figure as a snapshot, not a permanent fact. This article is for general information and is not legal or academic advice.

