How Turnitin detects AI writing

What happens to your file, in order

• The file is checked against the requirements: at least 300 words of prose, no more than 30,000, in English, Spanish or Japanese.
• The text is cut into segments of roughly five to ten sentences, and those segments overlap, so every sentence is scored with its neighbours around it.
• Each segment goes through the classifier, which returns a single number between 0 and 1.
• The sentence-level results are aggregated into the percentage your instructor sees, and the flagged sentences are highlighted in the report.
• If under 20% of the qualifying text came back as likely AI, no percentage is displayed. The report shows an asterisk instead.

The unit of judgement is a window, not your document

About five to ten sentences go into the classifier at a time. That is why one paragraph can come back highlighted while the paragraph beside it stays clean, and why deleting a sentence sometimes changes the highlighting on sentences you did not touch.

Because the windows overlap, each sentence is scored more than once, in different company each time. Turnitin says errors are more common in documents that mix human and machine writing, and specifically at the joins between the two. That is the mechanism behind it: a plain sentence sitting inside a smooth passage is read in the context of that passage.

What the model was trained on

Turnitin's architecture paper describes a transformer classifier in the BERT family, trained on two piles of text. One is human writing drawn from years of student papers already in Turnitin's archive. The other is text produced by GPT-family models.

So the question the classifier answers is narrower than the one people think it answers. Not 'did a machine write this' but 'does this look more like my human pile or my machine pile'. If your writing sits away from the human pile for reasons of your own, you start closer to the machine side before any assistant is involved. Writing carefully in a second language does that. So does a taught essay template, and so does a subject where the phrasing is conventional because precision requires it.

The 300 word floor, and what counts as prose

Turnitin's file requirements page sets hard boundaries, and a submission outside them gets no AI report at all rather than a low one.

• At least 300 words of prose in a long-form format, and no more than 30,000
• Written in English, Spanish or Japanese
• Uploaded as .docx, .pdf, .txt or .rtf, under 100 MB

Where the asterisk under 20% comes from

Turnitin stopped printing a number below 20% and prints an asterisk there instead. The reason it gave is the one that matters to you: the low band is where its false positives concentrate. A paper with a little flagged text is far more likely to be a mistake than a paper flagged throughout.

One consequence catches people out. An AI indicator with no number in it does not mean nothing was flagged. It can mean nothing crossed the line, or it can mean something did and stayed under 20%. If the distinction matters in your case, ask which of the two it was.

The two error rates Turnitin publishes

Turnitin reports a document-level false positive rate below 1% for documents where 20% or more of the text is flagged, and a sentence-level rate of around 4%. Both figures are the vendor's own. No one outside Turnitin can reproduce them, because the model is not available outside an institutional licence, and we have not measured it either.

The sentence figure is the one that reaches you, because highlighted sentences are what an instructor actually reads. Take an essay with 40 scoreable sentences. At a 4% sentence rate, between one and two highlighted sentences are expected to be wrong in a paper where nothing is wrong at all. A scatter of yellow across an honest essay is the predicted result, not a surprise.

Questions people ask about the number

What it cannot see

• How the text got into the document. A file upload carries no record of typing or pasting.
• Which assistant was used, or whether one was used at all.
• The editing you did. A generated draft rewritten by hand and an honest draft polished flat can score alike.
• Anything outside the file: your outline, your notes, your version history. Those stay yours to produce, and for most students they are the strongest evidence available.