The Turnitin AI Percentage Is a Share of Text, Not a Verdict

A Turnitin AI percentage is a share of the submitted prose, so 27 percent means the model read roughly a quarter of the qualifying text as likely machine-written. Turnitin's FAQ defines the figure as the amount of qualifying text its model determines was likely generated by AI or likely generated and modified by an AI paraphraser. Qualifying text means prose sentences only, and the figure is not necessarily a percentage of the whole submission ⁸.

Lists, bullet points and other short non-sentence structures are not analyzed ⁸. Nothing in that arithmetic estimates how the document was produced, and nothing in it is a probability that a person cheated.

Two numbers arrive in the same report. They don't measure the same thing.

The Similarity score is the percentage of matching text found when a submission is compared against Turnitin's collection of content ⁸. The AI writing detection percentage is the share of text its model predicts was generated or modified by AI writing tools ⁸. The two are completely independent and do not influence each other ⁸, which means a paper can read zero on one and high on the other. That is why the difference between the two scores is worth settling before any conversation about either. How the indicator's colors and states read, line by line, is covered on its own page. This page is about what the percentage decides and what it leaves open.

Getting any number at all has a length condition. Turnitin raised its floor from 150 words to 300 in May 2023, saying accuracy improves on longer text ¹. A submission with less than 300 words of prose comes back with no AI reading at all ⁸. That blank is a refusal to answer.

The asterisk warns that the reading under it is one Turnitin trusts less. Turnitin added it in May 2023 and gave the reason in the same post: below 20 percent the score is less reliable, and its chief product officer wrote there that false positives are more common in documents where the model detects less than 20 percent AI writing ¹. K-12 Dive reported the change at the time ⁴.

The display has since gone further. Turnitin's current FAQ says no score and no highlights are attributed in the 1 to 19 percent range. The indicator shows an asterisk with no percentage beside it ⁸.

A student who hears a number under 20 is hearing one the report no longer prints.

The error rate quoted in meetings carries a condition that rarely travels with it. Turnitin's document-level figure of under 1 percent applies only to documents its model reads as 20 percent or more AI writing ². Below that line, the post gives no rate.

Sentence highlighting has arithmetic of its own. Turnitin puts the rate at around 4 percent. Its May 2023 post defines that as the likelihood that a specific sentence is highlighted as AI-written when it is human-written ¹; the same post and a later one also put it as about 4 in every 100 highlighted sentences being human-written ¹². Those are two different quantities, and Turnitin's pages do not say which was measured.

Turnitin's FAQ sets no acceptable percentage. It says the figure should not be used as the sole basis for action or as a definitive grading measure ⁸, and Turnitin states that it makes no determination of misconduct, leaving the decision to the educator and institutional policy ³.

Twenty percent is a reporting line. Below it the indicator shows an asterisk and no highlights ⁸. Above it the published document-level error figure begins to apply ². Reading that line as a standard turns a note about the instrument's reliability into a disciplinary rule, and none of Turnitin's cited pages makes one ³⁸.

The band is crowded. Turnitin reported more than 200 million papers reviewed in the first year of the feature, of which over 22 million, about 11 percent, carried at least 20 percent AI writing by its own measure ⁵. A reading in that band puts a paper among millions of others, and that describes a population.

An institution that wants a number written into its rules meets that problem at full scale, which is the argument for leaving a percentage out of the policy.

The percentage can't say which tool produced a sentence or whether a grammar checker rewrote a clause. It can't say how long a paragraph took to write. Turnitin states that it makes no determination of misconduct even in the older territory of text similarity ³. The University of Pittsburgh's teaching center goes further: detectors are not reliable enough to act as proof ⁶.

The distance between a reading and a conclusion has a documented case. Marley Stevens, a student at the University of North Georgia, said she was put on academic probation for a year after a disciplinary hearing found she had cheated. She insisted she wrote the assignment herself, using only Grammarly's standard spell-checking and grammar features ⁷. Bloomberg Businessweek reported the case in 2024, in the same piece that records Turnitin declining to make its service available for testing ⁷. The dispute is over what the reading meant. A percentage can't settle that on its own.

How these tools behave on writers who learned English later is a separate question with its own evidence, gathered here: a detector flag on a non-native writer.

An allegation describes a person, and a percentage never gets past describing prose.

Ask the instructor for the AI report first. Turnitin's AI writing indicator and report are not visible to students, but an instructor can download the report as a PDF and share it ⁸. It highlights the text segments the model predicts were written or modified by an AI tool ⁸, and a list of sentences can be checked against a drafting record in a way a single figure never can.

Turnitin's guidance is to use the information to initiate a conversation, not to draw a conclusion ², and a drafting record is evidence a percentage does not contain. Version history in Google Docs or Word carries timestamps, and so do saved search results and interim files. Assembling that record gives that conversation something to check the highlighted sentences against.

What follows a flag is decided by the institution's academic integrity process, and knowing which stage a conversation has reached changes what belongs in a reply.

A second reading is worth more when the tool says how often it is wrong. Human's own in-house measurement, as of September 2026, 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. Those figures were measured on Human's own held-out sets and are not an independent audit; they sit on its research page. The public check at human.olive.is reads up to 1,200 words, one paper a day, with no account. Human's detector scores the text in memory and de-identifies any personal information for additional training. Like a Turnitin percentage, a reading there describes prose.