How plagiarism checkers work, and what the percentage means

The number at the top of a report is usually called the similarity index or similarity score. It is the share of your document that matches text somewhere in the checker's collection of sources. Nothing more.

It does not measure plagiarism. It cannot tell whether a match is a properly quoted definition, your reference list, a common phrase or a copied paragraph. It only knows that the same words exist somewhere else. Deciding what a match means is a human job: yours, your guide's, and sometimes a committee's.

Similarity indexthe headline number

The percentage of a document's text that matches sources in a checker's collection, after any exclusions the person running the check has set.

So 18 percent can be perfectly fine, and 5 percent can hide a serious problem. The number starts the conversation; it does not end it.

Checkers do not read the way people do. Vendors keep their exact methods to themselves, but the broad idea is well known: your document is broken into small overlapping pieces, and each piece is turned into a compact code, often called a fingerprint, that comes out the same whenever the same words appear.

1. Clean the text.Formatting, capital letters and often punctuation are set aside, so small changes do not hide a match.
2. Cut it into overlapping pieces.A sentence such as "the tablets were stored at room temperature" becomes a chain of short word runs that overlap by a word or two.
3. Turn each piece into a fingerprint.A short code is worked out for every piece. Identical pieces always give identical codes.
4. Look the fingerprints up.The codes are compared with those stored for a huge collection of documents. Where many codes match in a row, the checker marks a matching passage.

Because the pieces overlap, changing one word in a sentence rarely hides it: the pieces on either side still match. This is also why swapping a few words for synonyms is a poor disguise.

Each checker compares your work with its own collection. The large ones include:

• public web pages, including blogs, tutorials and course notes
• published journal articles and books; many publishers share their full text through Crossref Similarity Check, which is powered by iThenticate
• earlier student submissions stored by a college or by the vendor, where the institution allows it
• theses and dissertations in open repositories

Different tools hold different collections, so the same document can score differently on Turnitin, iThenticate and DrillBit, which many Indian universities use. That is normal. Compare reports from the same tool, run with the same settings.

A report usually has three parts: the overall percentage, your document with the matching passages highlighted in colour, and a list of sources, each with its own percentage and a colour that ties it to the highlights.

1. Look at the source list, not the total.One source with 12 percent means something very different from twelve sources with 1 percent each.
2. Open the biggest matches first.Read each highlighted passage next to the source it matches.
3. Sort every match.Is it a quotation you marked and cited, your reference list, a standard phrase, a name or title, or text you copied without credit?
4. Check the settings.See whether quotations, the bibliography and very small matches were excluded, as your rules allow.
5. Fix only what needs fixing.Quote and cite what should be quoted, rewrite what you copied, and leave common phrases alone.

Sai's 18 percent broke down like this. Six percent was his own reference list. Four percent was standard wording in his methods, such as the names of tests and equipment. Three percent was definitions he had quoted and cited correctly, and two percent was scattered matches of a few words each. The last 3 percent was a test procedure he had pasted from the Indian Pharmacopoeia with no quotation marks and no citation. Only that part was a real problem.

Plenty of matches are innocent. Most checkers let the person running the check exclude some kinds of match, and many rules expect it.

Zero percent is not the goal. A thesis with no matches at all may simply cite nothing. What matters is that every match is either legitimate or fixed. For the different kinds of copying a report can reveal, see types of plagiarism.

In India, the University Grants Commission set out plagiarism levels in the UGC (Promotion of Academic Integrity and Prevention of Plagiarism in Higher Educational Institutions) Regulations, 2018. For student theses and dissertations they work like this:

The regulations also list material to leave out of the calculation, such as quoted work with proper attribution, references and bibliography, and standard terms, symbols and equations. Many universities set their own stricter limits or procedures, so check your department's rules as well.

Sai's raw 18 percent would have put him in Level 1. With the reference list and quotations excluded, as the rules allow, and the pasted procedure fixed, he was comfortably inside Level 0.

A low number does not prove that work is original. Similarity checkers are good at one thing, finding matching text, and they struggle with anything that leaves no textual trace:

• a good paraphrase of someone else's idea with no citation
• text translated from another language
• borrowed ideas, arguments or structure
• images, figures and data tables, which some tools do not check at all
• sources that are not in the checker's collection

Examiners and reviewers fill these gaps by reading. They know the literature of their field, and a borrowed argument is often easy for them to spot even when the words are new.

AI-writing detectors are a separate kind of tool, and far less reliable. They guess from patterns in the writing instead of finding a source, and studies have found they can wrongly flag honest work, especially writing by people who use English as a second language. A detector's score alone should never be treated as proof.

If a report shows a real problem, fix it at the source.

• Put exact words in quotation marks and cite them.
• Rewrite copied passages from your own understanding, and cite the idea. Our guide to paraphrasing, summarising and quoting shows how.
• Replace long pasted procedures or standards with a short summary and a citation to the standard.
• Disclose any reuse of your own earlier work.
• Ask for the exclusions your rules allow, such as the reference list and quotations.

Never try to trick the checker. Word spinners, swapping letters for look-alike characters from other alphabets, hidden white text and pasting text as images are all forms of misconduct. Many checkers now flag unusual characters and hidden text, and even when a trick slips through, the strange writing it leaves behind is obvious to any examiner.

The easiest way to keep the number honest is to never lose track of your sources, which is the subject of how to avoid plagiarism. Tools can help with that part: Fiberarticle finds real papers, shows the passage behind each point it writes and formats citations in the style you need. It does not check plagiarism, so still run the checker your university provides.

Sai wrote a short summary of the pharmacopoeia procedure and cited it properly. The new report came back at 6 percent, almost all of it standard method wording, and Dr. Swathi signed the certificate the same afternoon.