A Turnitin AI Score Counts Flagged Sentences, Not Students

The percentage is a share of the paper's prose, not a probability about the student who submitted it. Turnitin defines it as the amount of qualifying text its model determines was likely generated by AI or likely modified by an AI paraphraser or bypasser. Qualifying text means prose sentences only. Lists and bullet points drop out, along with other non-sentence text, so the figure isn't necessarily a share of the whole submission ⁸.

Turnitin publishes its error rates in those terms: one figure for documents at 20 percent or more AI writing, another for a single highlighted sentence ¹. None of those figures is a record of how the paper was written.

The percentage isn't the similarity score, either. That figure counts matched text against Turnitin's index of sources. The AI indicator comes from a separate model with its own number, and one report can carry a low one beside a high one. People mix the two up often enough that the difference has a page of its own: similarity score against AI score.

Short work gets no reading at all. Turnitin raised the minimum from 150 words to 300 in 2023, reasoning that its detector evaluates longer text more accurately ³, and its current requirements still ask for at least 300 words of long-form prose ⁸. What comes back for a 280-word reflection? A gray indicator with no percentage ⁸. That's worth knowing before an assignment is built around the report.

Because Turnitin withholds the number there. In the current report no score or highlights are attributed in the 1 to 19 percent range. AI detected below the 20 percent threshold shows up as an asterisk (*%) with no percentage, a choice Turnitin says it makes to avoid false positives ⁸. A teacher looking at *% learns that the model found something. How much, and where, the report won't say.

The rule has a history. In 2023 the company first added an asterisk beside AI percentages under 20 percent, in its own words "to call attention to the fact that the score is less reliable" ³, and in the same post its chief product officer said a higher incidence of false positives sits in that band ³, which trade coverage repeated ⁴. Later the number itself came off the screen.

The sentence-level rate is a separate figure. By Turnitin's own count, roughly one highlighted sentence in every 25 may be human-written ¹. Take a paper with 50 highlighted sentences: about two of them may be.

The figure under 1 percent that gets quoted in meetings doesn't cover the asterisk band. Turnitin attaches it to documents with 20 percent or more AI writing ¹, and states no rate for the band below that line.

So an asterisk carries no error rate and no number to argue about.

Misconduct. Turnitin states that it does not make a determination of misconduct even in the space of text similarity, and that it supplies data for educators to decide under their own academic and institutional policies ². Its advice to instructors is to use the information to initiate a conversation, not to draw a conclusion ¹.

That leaves the threshold with the institution. It belongs there, and it's also where things tend to go wrong. A policy naming an acceptable percentage turns a model's estimate into a rule, when the vendor itself says the percentage should not be the sole basis for action ⁸. Turnitin's document-level claim is conditioned on documents at 20 percent or more AI writing ¹, so 20 is a boundary in a vendor's error analysis. It is not a line between honest and dishonest work. Administrators asking the same question will find it at why no AI percentage belongs in a policy.

What a flag can cost a student is documented. Bloomberg Businessweek reported that Marley Stevens, a student at the University of North Georgia, said she was placed on academic probation for a year after a disciplinary hearing, and that she wrote the assignment herself using only Grammarly's standard spell-checking and grammar features ⁵.

Common enough across Turnitin's own volume that the number alone sorts nobody. In the detector's first year Turnitin said over 200 million papers had been reviewed, and that over 22 million of them, roughly 11 percent, carried at least 20 percent AI writing by its own measure ⁶. One class's essays may not match that worldwide share, but at that rate a reading above 20 percent is not rare.

Volume turns a small error rate into a countable number of papers. Vanderbilt submitted 75,000 papers to Turnitin in 2022, set the company's claimed 1 percent against that total, and calculated that around 750 papers could have been incorrectly labeled in a single year ⁷. It switched the detector off in August 2023. Turnitin's own chief product officer wrote that behind each false positive is a real student ³.

Two papers reading 38 percent aren't the same case, either. One may carry a dozen consecutive highlighted sentences through a literature review, dense with definitions and summary. In another the highlights are scattered across a personal narrative and land on its plainest sentences. The percentage can't tell those apart. The highlighted sentences underneath it can, and that's the argument for opening the report instead of reading the indicator on the summary screen.

Open the full AI writing report. Its highlighted sentences are the only part of the reading that can be checked against anything. Turnitin says its wrong highlights are more common in papers that mix human and AI writing, at the seams between the two ¹. Read them beside the rest of the paper.

Four things are worth gathering before a student is contacted:

• The draft history, where the assignment produced one in a document editor or a learning platform.
• The sources the essay cites. Do they exist, and do they say what the essay claims?
• Earlier writing from the same student, read for register.
• The submission's word count, since anything under 300 words of prose never had a reading to begin with ⁸.

Then talk about the writing. Which paragraph was hardest? Which source got dropped, and what did the argument nearly say? Five minutes of that tells a teacher more than an indicator does, and it's the step the vendor itself recommends ¹. A guide to that conversation is how to talk to a student about an AI flag.

Students come to this number from the other side, after a grade or an email, and they type a different question. That page is what a Turnitin AI percentage means for a student.

A second reading is worth having when it comes from a tool that publishes its error bars, and it deserves the same care. Human reads a paper sentence by sentence and returns a verdict of Human, Mixed or AI; the public check takes up to 1,200 words, one paper a day, with no account, and a school or university address opens up to 50 papers a month free. Human's detector scores the text in memory and de-identifies any personal information for additional training. That gives a teacher holding only a Turnitin percentage something to set beside it. By Human's own in-house measurement on its own held-out sets, not an independent audit, its detector shows 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%, as of September 2026. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. The ceiling describes documents, not students, and a Human verdict is read the way the Turnitin report is, highlighted sentences first. The measurements and the run they were taken on are on Human's research page.