How Plagiarism Checkers Actually Work

A plagiarism checker never actually reads your paper against the internet. It cannot. Comparing your 2,000 words to billions of documents word by word would take longer than your degree. What it does instead is turn your text into a few hundred numbers, look those numbers up in an index, and only then fetch the handful of documents that shared any for a real comparison.

Once you know that, most of the checker's strange behavior stops being strange. Why it catches a sentence you reworded but misses a paragraph you translated. Why quotes match. Why two checkers disagree. All of it follows from how the numbers are made.

Step one: your text becomes overlapping chunks

The software first normalizes your document. Formatting goes, capitalization is flattened, spacing is standardized, and in most systems punctuation is stripped. What remains is a plain stream of words.

That stream is then cut into overlapping runs of consecutive words, usually somewhere between five and nine words long, called k-grams or shingles. Overlapping matters. A ten-word sentence does not become two chunks, it becomes six, each shifted one word from the last. This is why inserting a word in the middle of a copied sentence does not hide it: most of the surrounding chunks are untouched.

Step two: chunks become hashes, and most get thrown away

Each chunk is run through a hash function, turning it into a number. Comparing numbers is enormously faster than comparing strings, and the numbers are all the same size no matter how long the text is.

Keeping a hash for every chunk would still be too much data, so the system keeps only a sample, and which ones it keeps is the clever part. The standard approach, described in Schleimer, Wilkerson and Aiken's 2003 paper on document fingerprinting, slides a window across the sequence of hashes and keeps the smallest value in each window. Because the rule depends only on the values themselves and not on position, two documents sharing a passage will select the same fingerprints from that passage, even if the passage sits on page 2 of one document and page 40 of the other.

What survives is your document's fingerprint: a few hundred numbers standing in for thousands of words.

Step three: index lookup, then real comparison

Those fingerprints get looked up against an index built the same way from the checker's corpus. Documents sharing no fingerprints are eliminated instantly, which is nearly all of them. Only the small set that shares some gets fetched for a genuine text comparison, and that is what produces the highlighted passages and source links in your report.

So the percentage you see at the top is the last step in the process, not the first. The system located candidates statistically, then verified them literally.

What the corpus actually contains

A checker can only find what it has indexed, and this is where free tools and institutional tools genuinely differ. Turnitin's own database description puts its web index at 54 billion pages, crawled and archived over time so that deleting a page later does not remove it from the comparison set. Alongside that sit licensed publisher content and its repository of previously submitted student work.

That third one is the real asymmetry. No free checker can see the student paper repository, which is why a free tool cannot tell you whether your essay matches a paper submitted at another university last year. Any free checker claiming to replicate an institutional Turnitin score is claiming access it does not have.

What this design catches, and what slips through

Fingerprinting is excellent at some things and structurally blind to others:

• Caught easily: copied passages, lightly edited copying, and reordered material. Moving paragraphs around changes nothing, because fingerprint selection ignores position.
• Usually caught: synonym substitution, as long as enough consecutive words survive intact anywhere in the passage. This is why patchwriting so often surfaces despite looking rewritten.
• Often missed: heavy structural rewriting. If no run of five to nine consecutive words matches, there is no shared fingerprint and nothing to find, no matter how derivative the ideas are.
• Missed by design: translated text, ideas taken without wording, and anything in a source that was never indexed, including paywalled material and unpublished work.

Note what that last group has in common. A checker measures textual overlap. Plagiarism is about credit for ideas. These are different things, and no amount of engineering closes the gap between them.

Why your own paper can match itself

The most confusing report a student ever gets is the one showing 80% similarity against a single source that turns out to be their own earlier submission.

Nothing has gone wrong. If your institution's setup adds submissions to a repository, your proposal from last term is in the corpus, and your new chapter shares long runs of words with it. The system has no concept of authorship: it matched text, found text, and reported it. Whether reusing your own work is acceptable is a question for your supervisor and your institution's policy, and it is worth asking before the report exists rather than after.

The same mechanism explains the resubmission trap. Submit a draft to a repository-enabled assignment, then submit the polished version, and the second submission matches the first at close to 100%. Most instructors know to check for this. Not all of them do.

AI detection is a different machine

Worth separating, because they arrive in the same report and get discussed as one thing. Everything above describes matching: your words against documents that already exist. An AI indicator is doing something else entirely, estimating from statistical properties of the writing itself how likely a language model produced it, with no corpus lookup and no source to point at.

That is why the two numbers behave so differently. A similarity match can be verified, since there is a source document you can open and read. An AI estimate cannot be verified the same way, because there is nothing to compare against, only a probability. Two separate scores, two separate mechanisms, and a clean result on one says nothing about the other.

Why two checkers give you two different numbers

Three variables, all invisible to you:

Corpus. Different indexes contain different documents. A match one tool finds may simply not exist in another's index.

Chunk size. A system using five-word chunks flags shorter coincidences than one using nine. Neither is wrong, but the smaller setting produces higher scores and more noise from common phrasing.

Exclusion rules. Whether quotes, bibliographies, and short matches are filtered out changes the total dramatically before you ever see it. Two tools scanning one document can report very different percentages purely on filtering, which is worth remembering before you treat any single number as authoritative.

Reading a report with this in mind

Knowing the mechanism changes what you do with the output. Highlighted text means shared word sequences, nothing more, so your job is to sort the highlights into ones that need a citation and ones that are simply how the language works. Common phrases match because they are common. Reference entries match because citation styles are standardized.

The honest limit cuts the other way as well. When a report comes back clean, all it means is that the checker found no overlapping word sequences in its corpus. Originality is a different claim, and no matching engine can make it for you. The most a report can do is show you the source behind each highlight so the judgment stays yours. Ours marks exact matches red and reworded ones yellow, source link attached, instead of handing over a number and walking away.

The score is a measurement with a method behind it. Once you know the method, you can tell which parts of the measurement deserve your attention, and what the tool does with your text afterward becomes the more interesting question.