DetectorsFor teachers
Turnitin Scores Each Sentence of Prose, Then Averages Them Into One Percentage
Turnitin reads a submission in segments, scores each sentence of prose for whether it reads as human or machine written, and averages those scores into the single percentage an instructor sees [1]. Under 300 words, nothing is scored [2]. Between 1 and 19 percent no number and no highlights print; an asterisk does [8]. Turnitin itself says the product makes no determination of misconduct [1].
Olive Staff & OliveUpdated
How Does Turnitin Turn a Paper Into a Percentage?
The model reads a submission in segments, judges whether each stretch came from a person or from a machine, scores each sentence of prose, and averages those sentence scores into one prediction for the document 1. That average is the percentage on the report, and Turnitin's own account of the mechanism runs no longer than that.
Only prose sentences in long-form writing count; bullet points, tables and annotated bibliographies are not scored, so a document that mixes formats shows a percentage that does not match its highlights 8.
That average has a floor. Turnitin moved the minimum from 150 words to 300 in May 2023, on the reasoning that accuracy improves when the model has more text to read 2. A 200-word reading response or a short lab conclusion comes back with no AI percentage at all, and that silence says nothing whatever about the student who wrote it.
Inside the Similarity Report, the AI indicator is a separate control with its own percentage 2. Similarity and AI writing measure different things, so the two figures can differ on the same paper.
The two numbers are taken apart in similarity score versus AI score.
What Does the Asterisk Under 20 Percent Mean?
It marks the band where Turnitin says its own number is less reliable 2. In May 2023 the company added an asterisk to every AI percentage below 20 2. Its document-level claim, a false positive rate under 1 percent, is published only for documents containing more than 20 percent AI generated content 1.
Since 8 July 2024 the asterisk has replaced the number outright, with no score and no highlights attributed between 1 and 19 percent 8. Below that line the published record thins out. The University of Pittsburgh's Teaching Center records that Turnitin acknowledged a higher false positive rate in June 2023 and has never published a replacement estimate 7. A paper the model reads at 8 or 12 percent falls in that gap: an asterisk, no highlights, no vendor figure 8.
Highlighted sentences carry an error of their own. Turnitin puts the sentence-level false positive rate at around 4 percent, one highlighted sentence in every twenty-five, and reads that as about a 4 percent chance that a highlighted sentence is human written 3. It gives the rate without naming the documents it was measured on 3. None of that settles a single report. On a paper a student wrote alone, every highlight is the student's own prose, and nothing on the screen distinguishes that paper from one that was pasted in.
| Reading | What Turnitin publishes |
|---|---|
| More than 20 percent | Document false positive rate under 1 percent 1 |
| Between 1 and 19 percent | An asterisk, no score and no highlights 8 |
| One highlighted sentence | About 4 percent are human written 3 |
| Under 300 words | No score at all 2 |
What the number describes, and what it never describes, is worked through in what a Turnitin AI score means.
Which Submissions Does the Model Actually Cover?
Turnitin's report guide lists four supported languages, English, Spanish, Japanese and Arabic 8. Spanish arrived in September 2024, launched as a separate feature reporting an overall percentage the way the English version does 5. The newest layer reads English alone: its AI bypasser detection, announced in August 2025, carries a footnote saying the feature was trained and tested to support interactions in English only 4.
That footnote matters when a humanized paper arrives in a language other than English. Machine-written text run through a humanizer and handed in as a Spanish assignment falls outside what the bypasser layer was built and tested for, by the vendor's own statement 4. An AI writing percentage can still appear on the report, and Turnitin's report guide says its Spanish and Japanese detectors do not include bypasser detection at all 8. The evasion layer stops at the language line.
Whether any of these models read multilingual writers fairly is a separate question with its own evidence, taken up in detectors and non-native English writers.
Read the Report the Way Turnitin Tells Instructors To
Turnitin's own instruction is to use the information to initiate a conversation, not to draw a conclusion 3. The company states the same limit about the product as a whole, saying it makes no determination of misconduct and instead provides data for educators deciding under their own academic and institutional policies 1.
On a Tuesday afternoon with forty submissions and one report reading 34 percent, the useful next move is the drafting record. Where the draft was written in a tool that keeps version history, that history shows pace, pauses and pasted blocks, and it was there before any detector ran. How to read one is covered in version history as evidence.
Ask how the draft was built. A student who can talk through the sources, the paragraph that got cut and why the third section changed shape has answered a question the percentage never asked. The sequence after a flag, including what not to say first, is set out in when a paper gets flagged.
What the Report Cannot Tell a Teacher
It can't name the model that produced the text. Nor can it establish whether a student typed the words, or show the reasoning behind the number. Vanderbilt University gave that last gap as a reason for switching the detector off in August 2023, writing that Turnitin had published no detailed information about how it decides a piece of writing is AI generated 6.
Vanderbilt also did the arithmetic that a single report on a single desk hides. In 2022 the university submitted 75,000 papers, and at a 1 percent error rate roughly 750 of them could have been incorrectly labeled as containing AI writing 6. Vanderbilt applied Turnitin's rate to every submission, when Turnitin defines it on human-written documents and publishes it only for the over-20-percent band; the 750 is an illustration on an assumption, not a count.
Turnitin's chief product officer wrote in 2023 that "behind each false positive instance is a real student who may have put real effort into their original work" 2.
Pittsburgh's Teaching Center puts the conclusion plainly: detectors are not reliable enough to act as proof 7.
A rival detector's numbers sit under the same limit. Human publishes its figures, measured as of September 2026, as in-house measurements on its own held-out sets rather than as 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%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A ceiling measured on a held-out set stays a claim about that set, and no percentage from any vendor describes the paper on one teacher's desk.
Common questions
Does Turnitin score a paper under 300 words?
No. Turnitin raised the minimum from 150 words to 300 in May 2023, saying accuracy improves when the model has more text to read 2. A submission below that floor returns no AI writing percentage. A short reflection that draws no number is behaving exactly as designed.
Where does the AI score sit in the Similarity Report?
In its own indicator inside the Similarity Report, with a percentage separate from the similarity figure 2. Since 8 July 2024, between 1 and 19 percent no number and no highlights print: the indicator shows an asterisk 8. A high similarity score beside a low AI score is possible, and so is the reverse. The two measures answer different questions, and neither is a version of the other.
Does a Turnitin AI score say a student cheated?
No, and Turnitin says so first. The company states that it makes no determination of misconduct and provides data for educators to decide under their own academic and institutional policies 1. Its guidance to instructors is to use the information to initiate a conversation, not to draw a conclusion 3. The finding belongs to the instructor, and it rests on evidence the report does not hold.
Does the detector work the same way in every language?
No. Turnitin's report guide lists four supported languages, English, Spanish, Japanese and Arabic 8. Spanish AI writing detection launched in September 2024 as a separate feature reporting an overall percentage 5. Turnitin announced the AI bypasser layer in August 2025, and its footnote says the feature was trained and tested in English only 4. A paper in another language can still return a percentage, while Turnitin's report guide says its Spanish and Japanese detectors do not include bypasser detection at all 8.
How often is a highlighted sentence actually human writing?
Turnitin's own estimate is about one in twenty-five 3. The company puts its sentence-level false positive rate at around 4 percent, and gives that rate without naming the documents it was measured on 3. It is not a count on any one report. A paper the student wrote alone has no correct highlights at all, and the report cannot tell that paper apart from one that was pasted in.
References
- 1.Turnitin Originality | Amplify your academic integrity standard Turnitin, 2026. turnitin.comThe vendor's own description of the mechanism: a paper analyzed in segments, each sentence scored, the scores averaged into one prediction; the under-1-percent document figure for papers above 20 percent; and the statement that Turnitin makes no determination of misconduct.
- 2.AI writing detection update from Turnitin’s Chief Product Officer Turnitin, 2023. turnitin.comThe 300-word minimum, raised from 150 in May 2023, and the asterisk added to AI indicator percentages under 20 percent.
- 3.Understanding the false positive rate for sentences of our AI writing detection capability Turnitin, 2023. turnitin.comThe sentence-level false positive rate of about 4 percent, and the instruction to use the information to initiate a conversation rather than draw a conclusion.
- 4.Turnitin launches AI bypasser detection to combat AI misuse Turnitin, 2025. turnitin.comThe August 2025 bypasser detection release, whose footnote states the feature was trained and tested to support interactions in English only.
- 5.Turnitin introduces AI writing detection for Spanish submissions Turnitin, 2024. turnitin.comThe September 2024 launch of Spanish-language AI writing detection as a separate feature reporting an overall percentage.
- 6.Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector Vanderbilt University, Brightspace and Instructional Technology Support, 2023. vanderbilt.eduThe absence of published detail on how the detector reaches a verdict, and the arithmetic applying a 1 percent error rate to 75,000 papers.
- 7.Encouraging Academic Integrity University of Pittsburgh, University Center for Teaching and Learning, 2026. teaching.pitt.eduTurnitin's June 2023 acknowledgement of a higher false positive rate with no replacement estimate published, and the position that detectors are not reliable enough to act as proof.
- 8.Using the AI Writing Report Turnitin Guides, 2026. guides.turnitin.comThe current AI writing report: below the 20 percent threshold no score and no highlights are attributed and the indicator shows an asterisk with no percentage; reports generated before 8 July 2024 may still show a numerical score below 20.
8 sources, numbered by first appearance.
General guidance for teachers, administrators and students. What holds at one institution, on one assignment, may not transfer to another.
Human checks whether a piece of writing can be verified as written by a person, and shows the sentences behind the result. It does not report a probability that a person used AI, it does not check for plagiarism, and no number it produces stands for a student's honesty. A flag should be reason to look closer and not verified proof of cheating.