Direct answer
An AI plagiarism detector is not one tool but two checks running side by side: a similarity check that matches your text against existing sources, and an AI-writing check that estimates how much of your prose was machine-generated. The similarity side answers "did this text already exist somewhere?", while the AI side answers "does this text look like it came from a model?" Turnitin reports them as separate indicators, so a clean similarity score says nothing about your AI score, and vice versa [1]. Understanding that split is the whole game, because almost every confusing result students see comes from reading one number as if it were the other.
What Does an AI Plagiarism Detector Actually Detect?
The similarity half of the tool detects text overlap with existing documents. It compares your submission against a repository of previously submitted student papers, web pages, and publications, then highlights the passages that match and lists the sources they match against [2]. What it measures is string-level correspondence — the same words appearing in the same order somewhere else. It does not know whether you copied, quoted, paraphrased, or independently arrived at a common phrase, which is why the report shows you the matched source rather than pronouncing a verdict [2].
This is why a high similarity percentage is so frequently misread. A reference list, a block of properly cited quotations, a standard methodology paragraph, and a table of institutional definitions can all push the score up while the writing is entirely legitimate [2]. Turnitin's own guidance frames the Similarity Report as a reading aid: the number is a starting point, and the matched-source highlights are where the actual judgement happens [2]. Two students can submit the same score with completely different integrity situations.
The AI half detects something structurally different — statistical fingerprints of machine generation, not shared text. It is looking at how predictable your word choices are across long stretches of prose, which is a property of the writing process rather than of any source document [1]. That means a passage can score 0% similarity and still be flagged for AI, because nothing was copied; the text simply reads the way models tend to write. Conversely, heavily copied human prose can score high on similarity and low on AI. The two indicators are independent measurements of independent questions [1].
Practically, this changes how you should read your own report. If similarity is high, go to the highlighted matches and check whether attribution is adequate. If the AI indicator is high, source-hunting will not help you, because there may be no source to find — you need to look at the writing itself [1]. Treating the two as one blended "plagiarism score" is the single most common misreading, and it leads students to fix the wrong problem.
How Does an AI Plagiarism Detector Decide That Text Is AI-Generated?
Detection works on patterns across passages, not on individual sentences. The models assess long runs of qualifying prose and look for the kind of uniform, highly predictable phrasing that generation tends to produce, which is why a single suspicious sentence rarely moves the number on its own [3]. This also explains a detail that confuses many students: when the signal sits below the detector's confidence threshold, Turnitin displays *% instead of a precise figure, signalling a low-confidence result rather than a hard finding [1]. A *% is not a hidden high score — it is the tool telling you it does not have enough signal to commit.
The method has a known failure mode, and Turnitin states it plainly: false positives occur. Writing that is highly formulaic, heavily templated, or written by someone working in a second language can trip the pattern-matching even when every word is the author's own [3]. The same applies to text that has been run through grammar tools or paraphrased so thoroughly that the original voice is flattened out. Detection is therefore a probabilistic signal about style, not a forensic identification of who typed the words [3].
Because of that, the indicator is designed to be interpreted by a human with context, not to stand alone as proof of misconduct [1]. A percentage tells a reader that a stretch of prose resembles model output; it cannot tell them whether a model wrote it, whether the writer's style happens to resemble one, or whether an earlier draft was machine-assisted and then revised. Institutions that treat the number as a verdict are misusing the instrument, and the guidance around the tool is explicit that it is one input among several [1][3].
For students, the practical takeaway is that the detector is measuring a stylistic signature you can influence. If your prose is unusually uniform, or if whole sections were drafted by a model and lightly edited, the signal is genuinely there and the tool is reporting it accurately [3]. If your prose is your own but formulaic, the signal may be a false positive — and the fix is the same either way: change the writing, not the citation.
What Should You Do If Your Own Draft Is Flagged as AI-Generated?
Start by preserving your evidence of process. Keep dated drafts, version history, notes, reading lists, and outline files, because the strongest response to a flagged submission is a documented trail showing how the work developed [4]. This costs nothing to maintain and is far more persuasive than a retrospective explanation. If you have used a cloud editor, the revision history is already there — just make sure you know how to export it.
Next, rewrite the flagged passages in your own voice rather than arguing with the number. Turnitin's student-facing guidance points toward revision as the practical remedy: reworking the phrasing, adding specific sourced detail, and reintroducing the idiosyncrasies that make writing yours [4]. Vague, evenly-paced, generalising paragraphs are exactly what the detector responds to, so replacing them with concrete examples, named sources, and varied sentence structure does double duty — it lowers the signal and usually improves the grade.
Then talk to your instructor early and concretely. Because the indicator is meant to be one input in an integrity conversation rather than an automatic penalty [4], the useful move is to bring your drafts and ask how the result will be weighed under your institution's policy [1]. Approaching the conversation with evidence and a willingness to revise reads very differently from disputing the tool's output. Many institutions have explicit procedures for exactly this situation, and asking which one applies is a reasonable question.
Finally, check before you submit, not after. The reason flagged drafts feel catastrophic is that students discover the result at the same moment their instructor does. A pre-submission check turns an accusation into a revision task, because you see the same two indicators — similarity matches and the AI percentage — while there is still time to act on them [1][4]. That single change in sequencing removes most of the stress from the process.
If you would rather see those two indicators before your instructor does, turnitin0 gives you the same pair of documents — a Turnitin AI detection report and a similarity report — on your own draft, delivered in under 15 minutes in 98% of cases, without adding your file to Turnitin's student paper database. It is an independent service and is not affiliated with Turnitin, LLC, but the reports are built to match what you would see in your institution's system, so the numbers you revise against are the numbers that matter.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Does an AI plagiarism detector prove I used ChatGPT?
No. It reports a statistical resemblance to machine-generated prose, and Turnitin acknowledges that false positives occur, particularly with formulaic writing or work by non-native English writers [3]. The indicator is intended as one input in a human judgement, not as proof of authorship [1].
Why is my similarity score high when I did not copy anything?
Because the similarity check matches strings, not intent. Quotations, reference lists, standard methodological phrasing, and common definitions all register as matches against the repository, which is why the report shows you the matched sources to evaluate [2]. A high score with properly attributed sources is usually a formatting and citation question, not an integrity one.
What does *% mean on a Turnitin AI report?
It means the AI signal fell below the detector's confidence threshold, so Turnitin shows an asterisk instead of a precise percentage [1]. It is a low-confidence display, not a concealed high score.
Can I lower my AI score by paraphrasing?
Paraphrasing alone often does not help, because the underlying uniformity of the prose can survive a synonym swap. Rewriting flagged passages with concrete detail, named sources, and your own sentence rhythm is the approach recommended to students, and it targets the pattern the detector is actually measuring [4].
Should I check my draft before submitting it?
It is the sequencing that matters most. Seeing the similarity matches and AI percentage before submission converts a potential integrity dispute into an ordinary revision task while you still have time to act on it [4]. Turnitin0 provides both reports together for exactly that purpose.