Imagine this: a student submits a seminar paper. The plagiarism checker gives it the all-clear. And yet you cannot shake the feeling that you have seen these arguments, this sequence of ideas, before, only in different words.

This unease has a name: para-plagiarism. And it is no longer a marginal phenomenon.

What is the problem?

Is AI misuse as such still a fringe issue? No. The numbers are clear: 95% of people in higher education are convinced that AI is being misused at their institutions or that it is weakening students’ capacity for critical thinking. At the same time, 18% of undergraduate students in the UK openly admit to having submitted AI-generated text and that figure, of course, includes only those willing to say so. In addition, the available research suggests that up to 94% of fraudulently submitted texts remain entirely undetected.

Using online essay mills or drawing on a stockpile of previously validated work circulating among peers in order to pass off an existing paper as one’s own is nothing new. What has changed is this: today, a student does not merely copy a text. Instead, they ask AI to paraphrase it, or they use a so-called “humanizer” – an AI tool designed to make AI-generated text sound more human – and then submit the result. Thirty seconds. The underlying content, i.e. the intellectual property that was never independently developed remains unchanged. Research on machine-paraphrased plagiarism explicitly identifies the use of paraphrasing tools for concealment as a “serious threat” to academic integrity.

Why conventional tools fail here

The underlying problem is well documented at a technical level: traditional plagiarism detection measures lexical overlap. That worked as long as copying required real manual effort. That world no longer exists. A systematic review of 189 studies reaches a clear conclusion: lexical methods fail precisely where paraphrased plagiarism is concerned.

More troubling still, paraphrasing does not merely defeat plagiarism checkers. In the same move, it also bypasses AI detectors (which in any case should be used with caution), because the text has been paraphrased and therefore remains close enough to human writing to evade reliable detection. Two barriers, one tool, virtually no effort.

The central dilemma: all you have is the PDF

Where the assessment format allows it, close supervision can help. Alternatively, one can rely on intermediate steps that still bear visible signs of authentic human work. But in many assessment contexts, there is no such process evidence. No writing history. No earlier drafts. What remains is the final document and the quality of text analysis determines whether any reliable indications can be found at all.

This is precisely where a gap has existed until now.

Mentafy’s answer: “S³ – Semantic Source Search”

With our latest release, we close this gap. Our new algorithm, S³ – Semantic Source Search, extends conventional similarity checking with a semantic layer. Instead of comparing strings of words, S³ analyses meaning and argumentative structure across larger sections of text.

What this means in practice is straightforward: two sentences that express the same idea in entirely different wording can still be recognized as substantively related. When S³ identifies a semantically close passage, the system assigns the likely source and places it directly alongside the submitted text. You do not receive an abstract score, but a concrete side-by-side comparison that can serve as the basis for a discussion with the student.

The crucial point is this: while AI detection can, at best, provide indications rather than robust proof, and conventional plagiarism software often misses para-plagiarism, S³ identifies precisely those documents that served as the basis for the paraphrased submission.

See for yourself

S³ is now part of the Mentafy Academic Integrity Suite. If you would like to know what is really contained in your students’ submissions – beyond surface-level word matching – you can test para-plagiarism detection directly in your account.

Register free of charge now and try S³

Do you encounter borderline cases or disciplinary contexts in which para-plagiarism occurs particularly frequently? As always, we welcome your feedback — it helps us continue refining S³ as precisely as possible.