Your similarity score is not your plagiarism score.

Every August the same message arrives: a student has uploaded a draft, seen a number, and panicked. Sometimes it is 31%. Sometimes 6%, and the panic is about something else. Almost nobody has been told what the number counts.

So here it is, from the people who build the software and the UK universities that run it. The percentage measures matched text. It is not a measure of plagiarism, it is not a grade, and there is no safe number.

What Turnitin actually does

Turnitin compares your submission against a database — web pages, published articles, other students' submitted work — highlights every passage matching something it already holds, and reports how much of your document is highlighted.

That is the whole operation. It makes no judgement about whether a match is legitimate. The London School of Economics puts it in a section heading of its own staff policy: “Turnitin is not a plagiarism checker.” It is “a text-matching tool, not a plagiarism detection tool” which “cannot determine what those matches mean”. Strathclyde is equally blunt: without further investigation the report and score “must not be considered as evidence that plagiarism or collusion has or has not taken place”. Oxford Brookes tells students that “staff marking assignments are the detectors of plagiarism”.

Why 24% can be completely clean

Turnitin's own student guide asks “What is a reasonable Similarity Score?” and answers: “There is no fixed number to receive as a score in your Turnitin Similarity Report.” The reason is that a great deal of properly-done academic writing matches other text by design. The University of Reading's staff guide lists the categories of match that are routinely acceptable, and reading it dissolves most of the mystery:

• Quotations. An attributed direct quotation is a match by definition. You wanted it identical.
• Your references and bibliography. Other people cited the same works in the same style. Fifty Harvard entries will match.
• Matching formats — including the assignment title everyone on your module typed at the top of the page.
• Tables and charts reproducing shared or published data.
• Appendices, where cohorts reproduce the same instrument or standard.
• Common phrases and subject terminology. “A semi-structured interview schedule was developed” was never original prose.

A long, carefully-quoted literature review can reach twenty-something per cent out of nothing but good practice. Reading's colour bands put everything from 1% to 24% in the most common category. Not a licence — context.

Why 9% can be a serious problem

Now the other direction, which is less reassuring and therefore less written about. Reading's guide states plainly that “even a 1% score could potentially be plagiarised”, and that up to a quarter of a document inside the ordinary band could still have been copied without referencing.

A 9% score made up of three two-page passages lifted from one source, each reworded just enough to break the match and none of them cited, is a misconduct case. A 24% score made up of attributed quotations and a bibliography is a well-referenced essay. The percentage does not distinguish between them. A marker looking at where the matches fall does.

Oxford Brookes makes the sharpest version of the point: low scores do not indicate the absence of plagiarism, because essay-writing companies “are known to produce texts with 0% match”. Bought work is original text. It is also the most serious offence in the regulations, and the number cannot see it at all.

What the report does not see

Reading's guide lists the blind spots: images, drawings, diagrams and plans; printed books and journals not in the database; translated foreign-language works; and password-protected web content. If a low score has reassured you, that reassurance may simply mean the source was not indexed.

The actual trap: paraphrasing without citing

This is the one that catches honest students, and especially students writing in a second language under deadline. You find a passage that says the thing better than you can. You rewrite it in your own words, carefully, because you have been told that is what paraphrasing is. You move on without adding the citation, because it no longer looks like the original.

Two things are true at once. Reading's guide notes that Turnitin will often still highlight paraphrased text even when words have been changed, so the rewrite does not reliably hide anything. And, much more important: the offence was never about the words. Taking someone's idea, finding, argument or structure and presenting it without attribution is plagiarism whether or not a single phrase survives. The citation is what makes it legitimate, not the rewording.

The AI indicator is a different number, and a weaker one

Many UK departments now also see an AI-writing percentage. It deserves far more caution than the similarity score, and the clearest statements to that effect come from Turnitin itself.

In June 2023 Turnitin's Chief Product Officer, Annie Chechitelli, published the company's measured false-positive rates: under 1% at document level for documents containing 20% or more AI writing, and around 4% at sentence level — roughly one in twenty-five highlighted sentences may be human-written. The same post tells educators there is “no ‘right’ or ‘target’ score with the AI writing indicator, just like with a Similarity Score”, and to “use the information to initiate a conversation, not to draw a conclusion”.

Independent research found the error falls unevenly. Liang and colleagues, in Patterns in 2023, tested seven widely-used GPT detectors and found that over half of a set of TOEFL essays by non-native English speakers were misclassified as AI-generated, while accuracy on native-speaker essays was near perfect. Their explanation is uncomfortable but simple: detectors key on low lexical variety, and so does second-language academic prose. If you are an international student writing carefully in formal English, a detector is likelier to flag you for writing the way you were taught to.

What an originality audit is — and is not

We run originality audits as a mentorship service, and the phrase is used by other people to mean something we will not do, so here is the distinction.

An audit means a mentor reads your report with you and works out what each match is: an attributed quotation, your bibliography, a standard phrase, over-reliance on one source, or a paraphrase missing its citation. You learn to tell the difference and you make the changes yourself.

It does not mean we rewrite your text to move a number. “Plagiarism removal” services do exactly that, and what they produce is a document you cannot account for in front of your marker. Our academic integrity policy sets out the line in full.

The short version

1. The percentage counts matched text. Nothing else.
2. There is no acceptable number. If your department publishes a local threshold, use theirs, not a figure from a blog.
3. Quotations and references inflate it legitimately; unattributed paraphrase deflates it illegitimately.
4. The AI indicator is probabilistic, has a measurable false-positive rate, and falls hardest on second-language writers.
5. Your dated drafts are your best evidence. Keep them.

About this post

Related guidance

Sources

Primary sources and UK institutional guidance only. No competing commercial page is cited or paraphrased anywhere in this post. Every external link was checked on 9 October 2026.