The score arrived at 23 percent, and the student sat looking at it for a long moment before opening the details. She had written every word of that essay herself, in three careful sittings, and now a website was informing her that nearly a quarter of it "matched other sources."
The report, when she finally read it, told a quieter story. Most of the flagged text was a properly cited block quotation. The rest was her reference list and the phrase "participants were randomly assigned to two groups," which appears, near enough, in every psychology paper ever written. Nothing in the report was plagiarism. The number had simply aggregated everything the machine could see into one alarming figure.
Understanding what the machine actually does turns that panic into information. The mechanics are simpler than the marketing suggests, and the gaps matter more than the matches.
Fingerprinting: how a machine reads a sentence
A checker does not read. It slices your text into overlapping word windows, usually called n-grams or shingles, and treats each window as a fingerprint. In practice, each window gets hashed, meaning converted into a short code that comes out identical whenever the window does. The checker then asks its database one question, over and over: have I seen this fingerprint before?
Window size is a dial. Short windows, three or four words, catch more and false-alarm more, because ordinary language collides constantly. Long windows, eight or twelve, demand longer verbatim runs before flagging anything. Two checkers with different dials return different scores on identical text, which is the first reason the numbers disagree across tools.
The database matters more than the algorithm
The shingling math is decades old and effectively solved. What separates one checker from another is the library behind it. Commercial services index crawled web pages, licensed journal collections, and archives of previously submitted student papers, and every one of those collections has edges.
The consequences follow directly. A source absent from the index is invisible, no matter how much of it you copied. An essay can score clean against the web while matching a syllabus PDF the crawler never reached, or a paper behind a paywall the service never licensed. The same text returns different scores on different platforms because no two libraries are the same. When a checker advertises superiority, it is mostly advertising its shelves.
Why false positives are structural, not rare
The 23 percent student met the three main culprits, and they are structural, not bad luck. Quotations match because the fingerprint cannot see your citation; whether you marked and attributed the words is invisible to the machine. Methods sections match because whole disciplines share stock sentences; "the data were analyzed using" belongs to the genre. Reference lists match because every entry is, by design, a copied string.
| What lights up | Why it fires | What to do |
|---|---|---|
| Cited quotations | Fingerprints match; citation status is invisible | Confirm the quote marks exist, then relax |
| Methods phrasing | Stock sentences belong to the genre | Nothing, unless your target venue says otherwise |
| Bibliography entries | Every citation copies its source | Nothing; some tools exclude references automatically |
| Short common runs | Ordinary language collides | Ignore overlaps of one or two ordinary words |
| Genuine overlap | You leaned too hard on a source | Rewrite it, or quote and cite it |
The professional way to read a report is to skip the headline number, read the matches, and sort them into "accounted for" and "needs work." A similarity score behaves like a readability score in this respect: it measures the surface of the text, not its meaning or its merit.
The privacy problem with free checkers
Those databases are expensive, and with free tools someone other than you is paying. Some free services fund themselves with advertising and partnerships. Others reserve rights over what you paste: the right to store it, and the right to add it to the comparison corpus their product runs on. For an essay that will be published under your name anyway, the exposure is modest. For an unpublished novel chapter, an embargoed study, a legal draft, or a confidential client report, pasting it into an unknown free checker hands a copy to a company whose business is keeping text, possibly with a license attached.
Before uploading unpublished work anywhere, answer three questions: does the processing happen locally or on a server, is the text stored afterward, and does it join the database? If the answers are not easy to find, assume the uncomfortable versions.
What our checker can and cannot do
Honesty time, because an article like this should not exempt its own publisher. Our plagiarism checker runs entirely in your browser. It compares your text against the sources you supply, and only those: paste in the documents, and it checks against them. Nothing is uploaded, because there is no server involved, and there is no web index behind it.
That design is a set of trade-offs, stated plainly. It cannot tell you whether your text resembles anything you did not provide, and it cannot browse the world's published output on your behalf. What it does well is the most useful check of all, the one you should be running anyway: comparing a draft against the actual sources it was built on, auditing a contractor's delivery against the reference material it was supposed to transform, or catching your own recycled text before an editor does, self-plagiarism being its own quiet hazard. For quick two-document comparisons, the similarity checker does the same job with less ceremony.
Frequently Asked Questions
Can a plagiarism checker detect paraphrasing?
Only the lazy kind. Synonym-swapped text leaves long matching runs, so patchwriting lights up immediately. A genuine restatement, restructured and re-argued, produces almost no matching windows, and idea-level borrowing produces none at all. Paraphrasing properly is the skill the tool cannot supply for you.
Why did two checkers give me different scores?
Different databases, different window sizes, different handling of quotations and references. The score is as much a property of the tool as of the text. Pick one tool and compare your work against itself over time.
Does a high similarity score mean I plagiarized?
Not by itself. Quotations, methods boilerplate, and reference lists can push a report to startling percentages with nothing wrong in it. Read what matched, and why; the number alone is a headline, not a finding.
Are free checkers safe for unpublished work?
Some are, and some store submissions while reserving rights over them, and the terms change without notice. For anything confidential or unpublished, the safe rule is local processing in your own browser, or no check at all.
Can these tools detect AI writing?
Similarity checkers cannot; that is a different instrument entirely. AI detectors exist but are unreliable in both directions, flagging human writing and passing machine writing often enough that institutions generally treat their output as a prompt for a conversation rather than as proof. AI writing and integrity is its own subject.
Read the report, not the number
The percentage is a table of contents, not a verdict. Next time you run a check, go straight to the highlighted passages and sort them into two piles: matches you can account for, and matches you cannot. The second pile is the only one that matters, and it is nearly always far smaller than the number implied.