Plagiarism detection is the use of software, and the human review that follows it, to find text that has been copied or closely reworded from another source without credit. The software does not decide whether anyone has plagiarised. It compares a document with a large index of other documents, highlights overlapping passages and reports how much of the text matches. The technique exists because copying has become trivially cheap: a page of text can be lifted, lightly reworded and republished in seconds, and no human reader can check it against billions of other pages.

The term covers three markets that share the same underlying mathematics. Universities and schools screen student work, most often through Turnitin. Academic publishers screen manuscripts before peer review. Web publishers and search engines look for copies of articles across the open web, where scraped, spun and AI-rewritten copies compete with originals for traffic and advertising revenue.

How matching works

Most text-matching systems follow the same sequence, whatever the brand on the report.

  1. Normalisation. The submitted text is stripped of formatting, converted to lower case and often cleared of punctuation and very common words, so that trivial differences do not hide a match.
  2. Shingling. The text is cut into overlapping runs of words or characters, known as k-grams or shingles. A five-word shingle of "the cat sat on the mat" starts with "the cat sat on the".
  3. Fingerprinting. Each shingle is converted into a number by a hash function. Systems keep only a sample. The best-known sampling method, winnowing, was published by Saul Schleimer, Daniel Wilkerson and Alex Aiken at the ACM SIGMOD conference in 2003. It keeps the smallest hash in each sliding window, which guarantees that any shared passage above a set length produces at least one shared fingerprint. The authors reported performance within 33% of a proven lower bound and described its use in MOSS, a code-similarity service widely used in programming courses.
  4. Lookup. The fingerprints are compared against an index built the same way from millions or billions of other documents.
  5. Alignment and reporting. Candidate sources are compared passage by passage, and the matching text is highlighted in a report.

Turnitin's Similarity Report is the most widely seen output. According to the company's product page, submissions are compared against student papers, premium publications and current and archived web pages spanning more than 20 years, in 170 languages. The report gives a percentage score for matching text. Turnitin states that "there is no 'right' or 'wrong' number" and that the tool flags possible plagiarism rather than deciding whether it occurred. Quotes count unless excluded. A flags panel marks replaced or hidden characters used to break matching.

Copyscape works differently. It holds no large index of its own. According to an evaluation by HTW Berlin, it relies on Google services to search the web. Its Copysentry product monitors the web and emails alerts when new copies appear, and Copyscape Premium lets publishers check the originality of content they buy from freelancers.

Search engines run a related process at much larger scale, mainly to remove duplicates rather than to assign blame. In a 2007 paper at the World Wide Web conference, Google engineers Gurmeet Singh Manku and Arvind Jain, with Anish Das Sarma of Stanford, described using 64-bit simhash fingerprints, derived from Moses Charikar's 2002 work, across 8 billion web pages. Two pages were treated as near-duplicates if their fingerprints differed in three or fewer bit positions, a threshold the authors chose because precision and recall were both near 0.75 in their tests. The paper preferred it to Andrei Broder's 1997 shingling technique, which needed about 24 bytes per fingerprint.

Who operates it

On the education side, Turnitin served more than 16,000 institutions and over 71 million students as of 2025, according to its Wikipedia entry, and its archive held 1.9 billion student submissions as of June 2025.

Academic publishing runs on a shared database. Crossref, the not-for-profit that registers digital object identifiers, launched CrossCheck on June 19, 2008 with iParadigms, Turnitin's then parent company, and 11 founding publishers including Elsevier, IEEE, Oxford University Press and Wiley Blackwell. Its database was set to cover more than 20 million articles. The service, now called Similarity Check, runs on Turnitin's iThenticate product.

Web publishers check commissioned work before publication and hunt for copies after it, while search engines decide which version of duplicated text to show.

Origin and evolution

Turnitin traces its roots to 1994 at the University of California, Berkeley, where graduate student John Barrie built an online peer-review system. Barrie and three colleagues formed iParadigms in 1996, the service was introduced in 1998, the precursor site Plagiarism.org opened publicly in April 1999 and Turnitin.com launched more widely in early 2000.

Web-scale copy detection developed in parallel, from Broder's 1997 shingling work to winnowing in 2003 and simhash in 2007. Copyscape, from Gibraltar-based Indigo Stream Technologies, co-founded by Gideon Greenspan in 2003, launched in 2004.

Consolidation came later. Advance Publications bought Turnitin for $1.75 billion in 2019. Turnitin had already bought rival VeriCite in 2018, and went on to acquire Unicheck in 2020 and Ouriginal, formed from Urkund and PlagScan, in 2021. In early 2023 it added a separate AI writing indicator, a different technology from text matching.

Why it matters for the marketing community

For publishers, plagiarism detection is a revenue question. Scraped copies of an article can rank alongside the original and carry someone else's ads. Google's spam policies, last updated on August 28, 2026, define scraping as "the practice of taking content from other sites, often through automated means", and list examples including copying content and modifying it "only slightly (for example, by substituting synonyms)". Since March 5, 2024, the scaled content abuse policy has also covered mass-produced pages built from feeds, search results or other sites, including through "automated transformations like synonymizing, translating, or other obfuscation techniques". The same policy page states that a significant volume of valid copyright removal requests against a site can be used to demote other content from that site.

Generative AI has widened the dispute. In December 2025, Inspired Taste co-founder Adam Gallagher complained that recipe sites were publishing "full plagiarized recipes riddled with errors" built from his site's photos and text. When CNN sued Perplexity on May 28, 2026, the complaint alleged that outputs were "identical or substantially similar to CNN's content", and cited verbatim reproductions of named articles. In both cases the evidence came from comparing outputs with originals, the core task of plagiarism detection.

Limitations and disputes

The tools miss a great deal. A 2020 study led by Tomáš Foltýnek and Debora Weber-Wulff, run by researchers in seven countries, tested 15 web-based text-matching systems with material in eight languages. The authors found that the systems identified some plagiarised content but not all of it, flagged some legitimate material and described the results as "sobering". Copyscape, in the HTW Berlin profile from the same programme, found more than half of the similarities in manually paraphrased text but no multi-sentence similarities in translated text.

Machine paraphrasing is a specific weakness. In a paper presented at iConference 2022, Jan Philip Wahle and colleagues fed Wikipedia passages through the SpinBot rewriting tool. Turnitin flagged an average of 98.7% overlap on the originals and 30.2% on the spun versions. PlagScan fell from 65% to 0.5%. A trained language model classifier beat both tools and human evaluators.

High scores also mislead. A 2009 Texas Tech study found that many matches were jargon or standard topic phrases rather than copying. Students at McGill University in 2004 and 2005 won the right to have work graded without Turnitin, and Mount Saint Vincent University banned submissions in 2006. In A.V. v. iParadigms, US courts ruled in 2008 and on appeal in 2009 that archiving student papers was fair use. The Markup reported in June 2025 that per-student fees ranged from $1.79 to $6.50.

On the web, the hardest question is who copied whom. Matching shows that two texts overlap, not which came first. That gap can be exploited. In March 2026, a takedown request alleged that a Press Gazette investigation copied a 2024 Verge article "word for word"; Press Gazette said its reporting was original, and Google reinstated the articles within 48 hours.

Not to be confused with

  • AI detection estimates whether text was produced by a language model. It looks for statistical patterns in the writing itself, not for a matching source, so original machine-written text can pass a plagiarism check and fail an AI check.
  • Duplicate content is a search term for identical or near-identical pages, often on the same site, handled through canonical tags. It is usually a technical issue, not misconduct.
  • Copyright infringement is a legal finding. Plagiarism can occur without infringement, for example by uncredited copying of public-domain text, and infringement claims run through processes such as the DMCA.
  • False positive rate is the error measure applied to any detector, including these tools. A similarity score is not itself an error rate.

Recent developments

Google's guidance on AI-generated content changed on October 1, 2026, when the company told sites to "manually factcheck and review all AI-generated content" before publishing. The page cites rater guidance against content with "little to no originality" but does not mention plagiarism.

Publishers are building their own evidence. A group of 31 UK sites organised by the Movement for an Open Web began billing AI companies £500 per article in June 2026, documenting use by questioning chatbots and recording their output. In July 2026, Inspired Taste paid home cooks $2,000 each to test recipes as part of its campaign against AI versions of its work.

Timeline

  • 1994 - John Barrie builds the peer-review system at UC Berkeley that becomes Turnitin.
  • 1996 - iParadigms is formed to commercialise the software.
  • 1997 - Andrei Broder publishes shingle-based document resemblance.
  • 1998 - The Turnitin service is introduced.
  • April 1999 - Plagiarism.org opens publicly.
  • 2000 - Turnitin.com launches more widely.
  • 2002 - Moses Charikar publishes the similarity technique behind simhash.
  • 2003 - Winnowing is presented at ACM SIGMOD; Indigo Stream Technologies is co-founded.
  • 2004 - Copyscape launches; McGill student Jesse Rosenfeld wins grading without Turnitin.
  • 2006 - Mount Saint Vincent University bans Turnitin submissions.
  • May 2007 - Google engineers present simhash near-duplicate detection on 8 billion pages.
  • June 19, 2008 - Crossref launches CrossCheck with iParadigms and 11 publishers.
  • 2009 - The Fourth Circuit affirms fair use in A.V. v. iParadigms.
  • 2019 - Advance Publications buys Turnitin for $1.75 billion.
  • 2020 - Foltýnek, Weber-Wulff and colleagues publish their test of 15 tools.
  • 2021 - Turnitin acquires Ouriginal.
  • 2022 - Wahle and colleagues show machine paraphrase defeats commercial checkers.
  • Early 2023 - Turnitin adds an AI writing indicator.
  • March 5, 2024 - Google introduces the scaled content abuse policy.
  • December 2025 - Recipe creators publicly accuse AI systems of plagiarising recipes.
  • May 28, 2026 - CNN sues Perplexity over alleged verbatim copying.
  • June 2026 - 31 UK publishers start billing AI firms £500 per article.
  • August 28, 2026 - Google's spam policies page is last updated.
  • October 1, 2026 - Google adds manual fact-check language to its AI content guidance.

Summary

Who: Turnitin, owned by Advance Publications, dominates education; Crossref's Similarity Check serves academic publishers; Copyscape serves web publishers; search engines such as Google run their own duplicate detection and enforce spam policies against scraped content.

What: Software that breaks text into fingerprints, compares them with an index of other documents and reports overlapping passages for a human to judge. It measures similarity, not intent or originality.

When: Commercial tools appeared in the late 1990s, with Turnitin introduced in 1998 and Copyscape in 2004. Core fingerprinting methods date from 1997 to 2007, and generative AI has reopened the debate since 2023.

Where: In learning management systems, journal submission platforms, publishers' editorial workflows and the indexing systems of search engines, in up to 170 languages.

Why: Copying costs almost nothing at web scale. Detection protects academic credit, editorial standards and publisher revenue, but it misses paraphrased and translated text, cannot say who wrote first, and depends on human judgement to turn a percentage into a finding.