What is Plagiarism Checker Software?

The core mechanism is text matching, not reading comprehension. Turnitin's own documentation explains that a submitted file is compared against three broad collections: the current and archived internet, a subscription database of journals and periodicals, and its institutional repository of papers previously submitted by students [2]. The system breaks your text into searchable units and looks for the same sequences elsewhere. When it finds them, it flags the overlap and records the source it came from [2].

What this means practically is that the engine is sensitive to sequence, not to meaning. If you paraphrase a paragraph closely enough that long strings of words stay in the same order, the tool can still catch it; if you rewrite the idea in genuinely your own phrasing, it usually will not [2]. That is why paraphrasing tools that merely swap synonyms perform badly against modern checkers — the underlying word order survives.

The comparison is also purely mechanical in a second sense: the software has no way to know whether you copied deliberately, cited correctly, or simply wrote a common phrase. Every one of those situations produces a match [2]. This is the single most important thing to understand before you look at a score, because it explains why well-referenced academic writing routinely produces a non-zero number.

Finally, the reference collections matter as much as the algorithm. A checker that only searches the open web will miss paywalled journals and will completely miss essays sitting in a university's own paper archive — which is exactly where the most damaging matches tend to come from [2]. When you compare tools, the size and composition of the database is a more meaningful question than the interface.

The headline percentage is a ratio, not a judgement: it is the proportion of your document's words that matched a source, divided by the total word count [3]. Nothing in that arithmetic distinguishes a correctly quoted block from a copied one. Turnitin's guidance is explicit that there is no "ideal" percentage and that acceptable ranges differ by discipline and assignment type — a law essay built on statute quotations and a reflective portfolio are not measured on the same scale [3].

The number is also highly sensitive to settings. Excluding quotations, excluding the bibliography, and excluding matches below a small word threshold can each move the figure substantially, so two reports of the same document can legitimately disagree [3]. This is why comparing your score with a classmate's, or with a score from a different tool, tells you very little. What matters is the match list underneath the number.

That match list is where the real information lives. Each highlighted passage is tied to a specific source, and the productive workflow is to open every match and ask one question: is this properly attributed? A 28% score made of correctly cited quotations and standard methodology phrasing is a clean document; a 9% score containing one unattributed paragraph lifted from a source is a serious problem [3]. Instructors read reports this way, which is why the percentage alone almost never decides an academic integrity case.

The practical takeaway is to treat the report as a to-do list rather than a grade. Work through the matches, fix the ones that are genuinely unattributed, leave the ones that are properly cited, and re-check if the total still makes you uneasy [3].

Similarity checking and AI detection are two separate reports produced by two separate systems, and conflating them is the most common misunderstanding students bring to this topic. A similarity report answers "does this text appear somewhere else?" An AI writing report answers a different question: "does this text have statistical characteristics associated with generative models?" [4]. A document can score 0% similarity and still be flagged as AI-generated, and a heavily quoted document can score high on similarity with no AI signal at all.

Turnitin's AI writing detection reports the percentage of qualifying text — long-form prose above a minimum length — that its model classifies as AI-generated, and it deliberately withholds a precise figure when its confidence falls below a 20% threshold, displaying an asterisk instead [4]. That asterisk is not a pass and not a fail; it is the system saying the signal is weak. Understanding this convention matters, because students who see a symbol rather than a number often misread it in both directions.

The two reports are also reviewed separately by instructors, who apply different reasoning to each [4]. A similarity match points to a concrete source you can be shown; an AI score points to a statistical pattern that is inherently probabilistic and has been the subject of well-publicised false positives on human-written text. That asymmetry is worth knowing when you decide how much weight to place on either number before submission.

For anyone drafting with ChatGPT, Claude or Gemini, the practical implication is that running a similarity check alone leaves half the risk unmeasured [4]. You need to see both reports on your own draft to know what an instructor will actually be looking at.

Reading about how these reports are built is useful, but the numbers only become meaningful when they are your own. turnitin0 lets you upload a draft and see the actual Turnitin AI detection report and similarity report — the same two PDFs an instructor sees in the LMS — before anything is submitted, with turnaround typically under 15 minutes and no addition to the student paper repository.

※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary

Get Real Turnitin AI & Similarity Report