Turnitin AI Detection Accuracy Is Two Rates and a Warning

Two false-positive rates, and one scope condition that travels badly. The document-level rate is under 1%, and Turnitin's May 2023 update states it for documents with more than 20% AI writing ², restated in the June 2023 post as 20% or more ¹. The sentence-level rate is around 4%, meaning roughly one highlighted sentence in 25 may have been written by a person ¹.

Both rates are three years old. Both are the vendor's own measurement rather than an outside one.

Turnitin's current product page carries a separate pair of figures: 0.014 for English language learners against 0.013 for native English writers, on documents meeting its 300-word minimum ³, with no date attached. The two numbers sit close together, which is the point the page makes.

The most useful line on the page carrying the 4% rate is not a rate at all. Turnitin tells instructors: "So, use the information to initiate a conversation, not to draw a conclusion." ¹ That is the company describing the limit of its own product, in 2023, in its own words.

The asterisk covers the band where most papers land, and Turnitin says plainly that a score there is less reliable ². It was added to every AI score below 20% in May 2023. The same update raised the minimum document length from 150 words to 300, on the reasoning that accuracy improves with more text ². A paper barely past that 300-word floor, coming back under 20%, sits in the weakest part of the instrument on both counts.

That band is not a rare edge case. Turnitin reported that of 38.5 million submissions processed by 14 May 2023, 9.6% came back above 20% AI writing ². The other 90.4% came back at 20% or less by subtraction, and Turnitin does not say how much of that was read at exactly 0%, the one reading below 20% that still shows a number ¹¹. Since 8 July 2024 a reading above 0% and below 20% prints as an asterisk with no score and no highlights ¹¹.

On Turnitin's own 2023 split, the paper that lands on a desk on a Tuesday afternoon is far more likely to be one of the 90.4% at 20% or less than one reading 94% ². It shows an asterisk and no number, and a student is waiting outside.

What a percentage in that band supports, and what it does not, is worked through in what a Turnitin AI score actually means.

Applying any published rate to a term's worth of submissions is arithmetic, and it is done in the false-positive rate arithmetic.

Less than the question deserves, but not nothing. Two peer-reviewed groups included Turnitin on a bench and reached different conclusions, one about the class of tools and one about this product. Weber-Wulff and colleagues tested 12 public tools and two commercial systems, Turnitin and PlagiarismCheck, and concluded that the available detectors are neither accurate nor reliable and lean towards calling text human-written ⁸.

Walters ran 16 detectors over 126 undergraduate essays, 42 written by students and 84 generated by ChatGPT-3.5 and ChatGPT-4. Turnitin came out in the top three on all three sets, beside Copyleaks and Originality.ai ⁹. Two small benches, then: one class-level verdict and one product-level one.

Bloomberg Businessweek reported Turnitin's 4% sentence-level figure and that Turnitin declined to make the service available for the magazine's own testing ⁵.

Wider work has covered the class of tools rather than this one. In Patterns, a 2023 study ran seven detectors over 91 TOEFL essays written by non-native English speakers and reported an average false-positive rate of 61.3% ⁷. Turnitin was not among the seven. The failure mode turned up in every one of those seven tools, and it sits awkwardly beside the near-identical rates Turnitin reports for English language learners ³.

A teacher holding an essay by a multilingual student is weighing a vendor self-report against a peer-reviewed result on other tools. That trade-off is the subject of how detection policy handles non-native English writers.

The other kind of evidence is institutional, and it accumulated. Vanderbilt switched the detector off in August 2023, saying it does not consider AI detection software an effective tool ⁴. The teaching center at the University of Pittsburgh reached the same position: it backs no AI detection tool, and records that Turnitin acknowledged a higher false-positive rate in June 2023 without publishing a replacement figure ⁶.

Those decisions, and what the schools that made them wrote down, are collected in the universities that turned off Turnitin's AI detector.

Start by writing down which of the two rates applies to the paper in hand. A document score above 20% carries the under-1% claim ¹; a reading above 0% and below it carries the asterisk and no number ¹¹. Turnitin first stated the under-1% figure with no scope condition ¹⁰⁴, then limited it in May 2023 to documents at more than 20%, restated in June as 20% or more ²¹. Pittsburgh records no replacement figure for the range beneath ⁶.

The nearest thing on Turnitin's current page is an undated pair of document-level rates, 0.014 and 0.013, published for a different comparison and with no 20% scope stated ³. A single highlighted sentence carries the 4% rate, the looser of the two ¹.

• Check the length before the score. Under 300 words Turnitin returns no AI writing score, by design since May 2023 ². A paper a little over that floor is being judged on very little text.
• Read the highlights as a sample, not a verdict. At a 4% sentence rate, a report highlighting 25 sentences is expected to contain about one a person wrote ¹.
• Ask what the document looked like three drafts ago. Version history, research notes, marked-up sources and a bibliography that matches the argument are evidence a detector cannot see, and they are the evidence a student can actually produce.

Naming the set beside the rate is the habit this page argues for, and Human publishes its own numbers that way. As of September 2026, Human's own in-house measurement, on its own held-out sets rather than an independent audit, is 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%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Like every rate above, it describes documents and not students; the measurements and the run behind them are on Human's research page.

A percentage alone does not belong in a disciplinary file, and the vendor and the two campuses cited here agree on that much. Turnitin's guidance on the page carrying the 4% figure tells instructors to use the reading to start a conversation rather than to reach a conclusion ¹. The University of Pittsburgh states that AI detectors are not reliable enough to act as proof of a violation ⁶.

What replaces the threshold is a sequence. A reading prompts a closer look; the look is at drafts, sources and the argument the essay actually makes; the conversation with the student happens before anything is written into a file.

Every step in that order can be taken back easily, and an entry in a disciplinary file cannot.

Turnitin has also said, on its own blog, that behind each false positive is a real student ². That sentence is the honest reading of every number on this page: a published rate describes a pile of documents, and a disciplinary file describes one person. Nothing in a percentage bridges the distance between those two things.