What is semantic similarity detection – and why does it matter?

Beyond matching words

For many years, similarity checking has worked by looking for identical strings of words. That works well when text is copied and pasted. But when a source is paraphrased, translated or rewritten with AI, the wording changes while the meaning stays the same – and word-matching tools often find little or nothing. Semantic similarity detection is designed for exactly that situation. Instead of asking whether two passages use the same words, it asks whether they say the same thing.

How does it work?

The details vary between products, but semantic similarity detection usually involves the following:

• Splitting the text into passages, such as sentences or short paragraphs.
• Turning each passage into a representation of its meaning. A language model converts the text into a set of numbers – often called an embedding – so that passages with similar meaning end up close to each other, even if they share few words.
• Searching for close matches among passages from other sources, even in very large collections.
• Ranking the matches by how close they are in meaning, so that near-identical passages can be told apart from looser resemblances.
• Working across languages. Multilingual models recognise the same meaning in, say, Danish and English.

Why does it matter now?

The main reason is that rewriting has become effortless. Paraphrasing tools and generative AI can reword a source in seconds, and research has found that established text-matching tools miss a large share of machine-paraphrased text. A low similarity score from a word-matching tool therefore says less than it used to.

Yet copying from other students has not gone away. Despite generative AI, it remains one of the dominant forms of plagiarism – only now the copied text is more often reworded or masked. That is exactly where semantic detection makes the difference.

Language is the other reason. Many students read in one language and write in another, and translation is an easy way to hide a source. In multilingual Europe, comparing meaning rather than wording closes an obvious gap.

What semantic similarity detection cannot do

Semantic similarity detection still needs a source to compare with. If a text was written from scratch – by a student, a ghostwriter or an AI – there may be nothing to match. It is also not the same as AI detection: it finds borrowed meaning, not machine-written text.

Similar meaning is not misconduct either. Students who read the same texts and answer the same question will often express similar ideas, and standard definitions or common knowledge will produce matches. Because semantic tools also find looser resemblances, the marker’s judgement matters even more.

How does WISEflow approach it?

In WISEflow, semantic similarity detection is part of Originality, UNIwise’s own similarity service, built into the marking tool. Our understanding goes a little further than the technology alone:

• Meaning, not just words. Originality finds passages that have been paraphrased, restructured, translated or “AI-washed” – rewritten with AI to disguise the source – as well as text copied word for word.
• Clear categories for the marker. Each matching passage is marked as an exact, strong or potential match, so the marker can see at once where to look first.
• Built for European languages. Originality covers all major European languages.
• Institutions control the comparison. Each institution decides who it shares its submission data with, and how, and which external resources and archives work is compared against.
• The marker decides. The report supports academic judgement; it never replaces it.

What should institutions consider?

• Policy: make sure academic integrity rules cover paraphrasing, translation and AI rewriting – not only copying.
• Training: help markers understand and assess strong and potential matches.
• Students: explain that rewording or translating a source without acknowledging it is still plagiarism – and that it can be found.
• Process: in a misconduct case, document the source – not just the match.

Semantic similarity detection does not replace academic judgement. Used well, as part of a considered approach to academic integrity, it closes the gap that rewriting tools have opened.