A Turnitin Similarity Percentage Is Matched Text, Not an AI Reading

Turnitin's similarity percentage is the share of a submission's text that matched something in Turnitin's databases. The company's own description is plain: the score is the percentage of text in a submission that matched against other sources, and material a student quoted counts toward it ¹. Nothing in the number describes who wrote the paper or how.

Turnitin says a submission is compared against seven trillion matches in its database, a pool that takes in internet articles and scholarly publications as well as student papers ¹.

That is a text-to-text operation. A sentence generated by ChatGPT on Monday and never published anywhere often matches nothing, so it sits in the report near zero. A sentence lifted from a course reading can match when that reading is among the stored sources, whether the student typed it as a quotation or not.

Low numbers are ordinary. Boston University's Turnitin FAQ tells instructors that a low similarity score is not necessarily unusual or reason for concern, and that zero percent similarity on twenty papers out of hundreds is typical in many courses ⁵.

One familiar exception has nothing to do with copying. When an instructor allows resubmissions without the overwrite setting, a second upload is compared against the first, and the report can come back at 100 percent similarity against the student's own earlier draft ⁵.

Nothing was copied there. The 100 percent was an earlier version of the same file.

A cited quotation raises the similarity score because the match is made on the text, not on the citation beside it. Turnitin says that if students use any text from source material to support their work, such as a quote, it will be reflected in the score ¹. An instructor can trim the count. The report carries a match exclusion feature for citations and bibliographic information, and applying it is an assignment setting rather than a property of the paper ¹.

The same essay can return different percentages under two assignments' settings.

Turnitin's product page names no safe number. It says an institution may or may not have assigned a threshold, and adds that there is no right or wrong number to receive as a score ¹.

A literature review that quotes generously runs higher than a reflective essay that quotes nothing, and neither figure on its own is evidence about either student.

The useful question is never the percentage on its own. It is which sources the matches point to, how long the matched passages are, and whether they carry quotation marks and a reference. A single 22 percent match to one classmate's paper is a different situation from twenty-two matches of one percent each spread across a reference list. The second of those is what a bibliography looks like when every entry follows the style guide.

The AI percentage comes from a separate product. AI writing detection reaches a Turnitin Similarity customer through the Turnitin Originality add-on. Once it is enabled, the AI writing indicator is linked to the Similarity Report and shows the overall percentage of the document that AI writing tools may have generated ¹. One number reports what matched stored text ¹; the other reports an estimate ⁷.

Linked in the interface is not the same as produced by the same process, and the AI indicator carries conditions the similarity score does not.

The AI reading needs 300 words. Turnitin raised that minimum from 150 in May 2023, on the reasoning that its detector evaluates longer text more accurately ³. Below the line no AI percentage appears at all, while the similarity score still does. A short paper gets one number instead of two.

Since 8 July 2024, from 1 to 19 percent no score and no highlights are attributed and the indicator shows an asterisk, while 0 percent is still displayed; reports generated before that date may still show a number under 20 ⁷. The asterisk arrived in May 2023, in the Similarity Report, to call attention to the fact that the score is less reliable in that band ³. Turnitin's pages describe no such marking for the similarity score ¹.

In 2023 Turnitin put its sentence-level false positive rate at around 4 percent, a 4 percent likelihood that a specific highlighted sentence might be human-written ⁴. Its document-level figure from the same year, under 1 percent, applies only to documents the model reads as 20 percent or more AI writing ⁴. The similarity report lists the sources it matched, so a match can be checked against the source; the AI indicator points to no source ¹⁷.

What that second percentage counts, and what the asterisk changes about reading it, is a question of its own: what a Turnitin AI percentage means.

Plenty of reports carry no AI percentage at all. Boston University switched its AI score panel off on 3 July 2023, naming false positives and a lack of supporting evidence for the tool's efficacy ⁵. On a campus that has done the same, the similarity number is the only one a student will ever see.

A zero on one report says nothing at all about the other. Zero similarity means the text matched no stored source, which is exactly what freshly generated prose tends to do. Zero on the AI reading means the model identified no qualifying text as likely AI-generated ⁷, which says nothing about whether a passage was copied from a classmate. The two reports answer different questions and neither covers for the other.

A humanizer is built to change an AI reading; it does nothing to supply a missing citation: a humanizer is not a plagiarism remover.

Turnitin is direct about the limit of both reports. Its own post states that the company 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 academic and institutional policies ². The same company tells instructors to use the information to start a conversation rather than to draw a conclusion ⁴.

A third kind of reading belongs to the estimate column rather than the match column. Human reads a piece of writing and reports how much of it reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI. The public check takes up to 1,200 words, one paper a day, with no account. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Human's detector scores the text in memory and de-identifies any personal information for additional training. None of that compares the submission against stored sources, so it settles the similarity question no better than a similarity percentage settles the authorship one.

Neither report knows who typed the paper.

Schools running Blackboard meet a version of the same question under another product name. Anthology says SafeAssign compares submitted assignments against a set of academic papers to identify areas of overlap between the submission and existing works ⁶. That is a matching operation rather than a reading of authorship: what SafeAssign actually checks.

Start by asking which report the message is about. An email that says 38 percent could mean matched text, machine-written text, or a grade, and the three lead to completely different replies. Ask for the report itself rather than the figure. The similarity report lists the sources it matched, less any the instructor excluded ¹, and at 20 percent or above the AI report highlights the text it marked ⁷.

Then ask about the settings. Whether citations and bibliographic information were excluded changes the similarity percentage without changing a word of the paper ¹.

Keep the draft history. A Google Docs or Word version history can show a paper arriving in pieces across several days, with the outline written before the paragraphs and the citations added last. That is the kind of evidence a percentage is not.

What follows a flag is set by the institution rather than by the vendor, and the steps vary widely: what happens if Turnitin detects AI.

Put the request in writing. A short message asking which report produced the number, which sources or sentences it points to, and which exclusions were applied puts the question and the answer in the record.

The percentage is where that request starts.