Turnitin Says It Detects ChatGPT, With Four Caveats Attached

Turnitin's AI checker page says the tool helps educators identify when an AI writing tool "such as ChatGPT, text spinners, or AI bypassers" may have been used in a student's submission ¹. Read the verb. Identifying when a tool may have been used is the whole claim; the page never says which product wrote the text. What lands in the Similarity Report is a percentage of the document, carried on an AI indicator button ².

For a teacher opening that report between classes, that distinction separates a lead from a verdict. The percentage reads text and nothing else. Whether a student prompted anything comes from the draft history, the sources, and a conversation held before any accusation gets written down. A report that lands at four in the afternoon, with the class due at nine the next morning, is a prompt to look, and looking takes longer than reading a number. The percentage itself is a narrower claim than it looks.

A score describes a document, and a document is not a person.

Three hundred words. That's the floor. Turnitin raised its minimum from 150 words to 300 in May 2023 and tied the change to accuracy on longer text ². Anything shorter comes back with no AI indicator at all, one reason a teacher who expects a number on a discussion post or a lab conclusion finds the panel empty ⁸.

Percentages between 1 and 19 no longer print in the Similarity Report. The asterisk arrived in that same May 2023 update to mark those scores as less reliable ², and since 8 July 2024 it stands in place of the number: no score and no highlights are attributed between 1 and 19 percent ⁸.

Four published conditions travel with every number on the screen:

• Papers under 300 words: no AI indicator at all, since May 2023 ².
• Readings between 1 and 19 percent: since 8 July 2024, no score and no highlights, only an asterisk ⁸; before that, the number with an asterisk beside it ².
• Highlighted sentences: around 4 percent of them are human-written, by the company's own count ³.
• Any reading whatever: Turnitin makes no determination of misconduct, and says the decision rests with the educator and institutional policy ⁴.

Under 1 percent of documents and around 4 percent of highlighted sentences, on Turnitin's two dated rates from 2023, which answer different questions. The document-level rate was first stated without conditions in March 2023 ⁴, then restricted from May 2023 to documents the model reads as 20 percent or more AI writing ^{2 3}. The sentence-level rate is the one a teacher sees on screen: roughly one in every 25 highlighted sentences was written by a person ³.

A highlighted sentence is a reason to read that sentence, not a finding about it.

Pittsburgh's teaching center records that in June 2023 Turnitin acknowledged a higher false positive rate than it had first asserted, and that no replacement estimate has been disclosed ⁵.

Vanderbilt did the arithmetic on its own volume before announcing in August 2023 that it was disabling the detector. In 2022, 75,000 papers went to Turnitin. At a 1 percent false positive rate, Vanderbilt wrote, around 750 of them could have been wrongly labeled ⁶. The arithmetic applies the rate to every paper submitted, not only to the human-written ones. Read it as Vanderbilt's assumption, not as a measurement. The university said the decision would hold for the foreseeable future ⁶. Other campuses reached the same place, and the roll call is a story of its own.

Turnitin also published an English language learner comparison: 0.014 against 0.013 for native English writers, on documents meeting the 300-word requirement ¹. That is 1.4 and 1.3 percent, measured on a different population from the under-1-percent figure, which Turnitin limits to documents read as 20 percent or more AI writing ². The page gives the comparison no date.

Turnitin's instruction to instructors is one sentence: "use the information to initiate a conversation, not to draw a conclusion" ³. Elsewhere the company states that it does not make a determination of misconduct, and that the decision rests with the educator and institutional policy ⁴. Pittsburgh's teaching center goes further and calls detectors "not reliable enough to act as proof" ⁵.

What a teacher can hold up in a hearing is what the conversation produces. The work starts where the report stops.

• The document's version history, opened before the meeting rather than during it. A paper written in one paste reads differently from a paper written across eleven sessions.
• A five-minute conversation about sources and choices: which argument came first, what got cut, why the third paragraph changed. Questions asked out loud give a panel something to weigh that a percentage does not.
• Earlier writing from the same student in the same course, a baseline the teacher already has.
• The report itself read past the headline: the asterisk, the highlighted sentences, the length of the document ².

The stakes are documented. Marley Stevens, a University of North Georgia student, told Bloomberg Businessweek she was placed on academic probation for a year after a disciplinary hearing that followed a Turnitin flag. The only tool she said she used was Grammarly's standard spell-checking and grammar features ⁷. What the hearing weighed isn't reported, which is the point: a draft history and a conversation are things a student and a panel can both examine. Where editing help ends and machine writing begins is a line no published Turnitin rate draws, which is a question of its own.

A flag buys a second look. Nothing more.

The dated ones, and every detector, Turnitin's included, owes a teacher two: what it does to human writing it has never seen, and what it misses. Turnitin's published answers to the first are the under-1-percent document figure from 2023 ³ and an undated English language learner comparison on its checker page, 0.014 against 0.013, which is 1.4 and 1.3 percent ¹. Pittsburgh's teaching center says no replacement for the 2023 estimate has been disclosed ⁵.

Human, at human.olive.is, publishes both with a date. Measured as of September 2026, as its own in-house measurement on its own held-out sets rather than an independent audit: 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. And 0 misses in 481 wholly machine-written documents; the ceiling on that is 0.62%. The research page at human.olive.is/research/model-card names the model run beside each figure. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A reading from Human earns the same second look as a Turnitin reading in the asterisk band, a lead rather than a verdict.

The limits travel across products. Short text gives any detector less to read. Below 50 words Human declines to answer, and between 50 and 149 words a reading is weak. Turnitin's 300-word floor exists for the same reason ².

A school comparing detectors can ask each vendor for two things: a dated false positive rate, and the size of the set it was measured on.