With generative AI, the question becomes more urgent. A language model can rephrase existing content in seconds. A text that sounds new is therefore not automatically original!
Conversely, a text with a high similarity is not automatically plagiarism.
Plagiarism checks are usually carried out primarily at universities.
A plagiarism checker only compares your text with the content it knows
The software breaks down a text into sections and searches for identical or similar passages in its comparison sources. Depending on the provider, these include publicly accessible websites, specialist articles and scientific publications, documents in proprietary databases, and previously submitted papers. For students, it is important to note: a check of term papers only finds matches with the content the provider has access to.
What is not in the content cannot be found by the software. A checker that only knows public websites will not see a match with an unpublished paper in a closed university database. The size, currency and composition of the content therefore determine the quality of the result more than any advertised detection rate. The internet is only part of this content; closed collections remain excluded.
In the end, a report is generated. It highlights the affected passages, names the sources found, shows the match per source and summarises them as a percentage value.
The check proceeds in four steps
First, the software prepares the text. It removes formatting and special characters and breaks the text down into words, phrases, sentences or larger sections.
Then it compares these parts with its content. Simple methods look for identical word sequences, advanced ones also consider similar phrasing.
In the third step, it searches for non-verbatim borrowings. Depending on the system, it compares word similarities, sentence structure or the meaning of sentences, sometimes using AI-supported methods. The more a foreign text is rephrased, translated or newly composed, the less certain this detection becomes.
Finally, it creates the report. This forms the basis for your assessment and does not replace it.
The software reliably finds verbatim borrowings, but hardly rewritten ones
ie gut eine Übernahme erkannt wird, hängt von ihrer Form ab. Die Universitätsbibliothek der Freien Universität Berlin weist darauf hin, dass wörtliche Übernahmen deutlich besser erkannt werden als paraphrasierte.| Form | What happens | How well the software detects it |
|---|---|---|
| Complete plagiarism | An entire foreign text is presented as your own. | Usually well detected if the source is in the content |
| Copy-and-paste plagiarism | Individual passages are copied word-for-word. | Generally well detected |
| Partial plagiarism | Larger sections are copied and mixed with own text. | Often well detected |
| Paraphrase plagiarism | Foreign content is presented in own words without source. | Limited depending on system |
| Translation plagiarism | A foreign-language text is translated and presented as own. | Only with multilingual comparison |
| Structural plagiarism | The structure and argumentation of a foreign text are copied. | Difficult |
| Idea plagiarism | Foreign ideas or concepts appear without source. | Especially difficult |
The further a borrowing deviates from the original, the less the software finds it. These are exactly the forms most easily generated by a language model.
The Similarity Score measures similarity, not guilt
Most checkers summarise their result as a percentage value, the Similarity Score. It indicates the proportion of the text for which the software has found matches. Turnitin, one of the best-known providers, emphasises that this value is not a plagiarism assessment.
An example shows why. An article has 1,000 words; the software finds matches in 120 words, so the value is 12 percent. Among the highlighted passages can be correctly marked quotes, common phrases, technical terms and definitions, recurring text blocks, and only then unauthorised borrowings. The value does not indicate how large the last part is.
Conversely, a low value does not prove originality. A heavily rephrased or translated text can show few matches even though its content comes from a foreign source. There is no universally valid threshold at which a text is considered plagiarism.
The Similarity Score is a check indicator, not a judgement on the text.
AI texts require three separate checks
A language model rarely reproduces a passage word-for-word. It has processed information from many sources and rewrites it. A search for identical word sequences then finds little, even if the content is close to an original.
Three methods therefore answer three different questions. The classic plagiarism check searches for passages matching existing publications. The semantic similarity check searches for content-wise similar and rewritten passages; how well it can do this depends on the method and comparison data. The AI text detection estimates whether a text is likely written by a language model.
An AI detector is not a plagiarism checker, and a plagiarism check cannot determine whether a human or a model wrote a text. How reliable AI text detection is today is described in our article Detecting AI Texts. Those publishing AI-supported content therefore do not rely on a single check value but verify sources, factual accuracy and the text’s own contribution.
For SEO, your own added value counts, not a low similarity score
Google does not penalise every text that contains similar phrasing to another page. For identical or very similar content, Google selects a preferred address, the canonical URL. This is a matter of indexing, not quality; more in the glossary under Duplicate Content and Canonical Tag.
Google asks the quality question elsewhere. In its guidance on helpful content, Google asks whether a text that builds on other sources does not simply copy or rephrase them but offers clear added value and originality. This is the core of the Helpful Content System and closely related to E-E-A-T.
A low Similarity Score therefore does not guarantee a good page. A completely original text can be superficial, factually incorrect or miss the search intent. Some similarity due to technical terms or properly marked quotes does not harm. The checker complements quality control but does not replace an SEO audit.
Plagiarism is not automatically a copyright infringement
German law does not have a general definition of plagiarism. The term mainly comes from academia and is also used in journalism and marketing.
Copyright law protects personal intellectual creations, including sufficiently individual texts (§ 2 UrhG). Anyone who uses and publishes protected passages without permission can infringe copyright. Whether this is the case depends on the content taken, its protectability and the
When choosing, content, sources and data protection matter
The comparison content comes first. Ask whether the software only knows public websites or also scientific publications and other document collections, and how current they are.
Next, it matters whether the software only finds identical passages or also rephrased ones. Especially with AI-supported texts, this is the difference between a usable and a reassuring check.
A percentage value alone does not help you. A good report shows every affected passage and links to the source found so you can decide whether it is a quote, a technical term or a borrowing. For multilingual content, it also matters which languages the software actually compares.
If you upload unpublished articles, client texts or internal documents, check data protection. The key questions are whether the provider stores the texts, includes them in its comparison content, uses them for training AI models, and where it processes them.
Free checkers are sufficient for an initial assessment. They usually limit text length and number of checks, do not specify their comparison content exactly, do not provide detailed reports and detect rephrasing poorly. Paid solutions offer larger content and more detailed reports depending on the provider. They also only find what is in their content.
How to integrate the check into your editorial workflow
Plagiarism checking works best as one step among several, not as the last hurdle before publication.
- Write the article based on your own research and verifiable sources.
- Check foreign statements, figures and quotes for origin and accuracy.
- Carry out the plagiarism check and assign each match to its source.
- Evaluate each source: quote, common phrase or problematic borrowing.
- Add missing source references, rewrite problematic passages and strengthen the text’s own contribution.
- Finally, check factual accuracy, readability, search intent and structure.
For AI-supported texts, step 2 also includes the question of whether every statement is supported. Language models sometimes cite sources that do not exist.
Frequently asked questions about plagiarism checkers
What does a plagiarism checker do?
It compares a text with the sources in its content and highlights identical or similar passages. The result shows where there might be unauthorised borrowing; whether it actually is, you decide.
How accurate are plagiarism checkers?
It depends on the method, the comparison content and the type of borrowing. They detect verbatim passages more reliably than rephrased, translated or conceptually adopted content.
What does a Similarity Score of 20 percent mean?
The software has found matches in about one fifth of the text. This does not mean that part is plagiarised; it can include quotes, technical terms and common phrases.
Does a plagiarism checker detect AI-generated texts?
Not reliably. An AI text can be independently formulated yet still reproduce existing content. Whether a language model wrote a text is estimated by separate AI text detection methods.
Do I need a plagiarism checker for SEO?
It is not a prerequisite for good rankings. However, it helps to find problematic borrowings before a text is published, thus securing editorial quality.
Are there free plagiarism checkers?
Yes. Many providers offer free checks, usually with limits on text length, number of checks and features. Paid solutions are usually more suitable for regular professional checks.
Is every duplicate content plagiarism?
No. Duplicate content means identical or very similar content under different addresses, for example due to technical URL variants, product descriptions or deliberately reused texts. Plagiarism requires the adoption of someone else’s intellectual work without acknowledgement.
A text does not become good by having as little in common with other texts as possible, but by having something original to say.
This article is based on these sources
- Freie Universität Berlin, University Library: What is plagiarism? and What can anti-plagiarism software do?
- Turnitin: Explanations on the Similarity Score and plagiarism assessment
- Google Search Central: Creating helpful, reliable, people-first content
- Johannes Gutenberg University Mainz: FAQ on software-supported plagiarism checks
- Copyright Act, §§ 2 and 51 UrhG