Translate it, and it's gone. That is the promise behind one of the oldest cheating tricks in academic writing. Now generative AI has made it effortless. Take a source published in another language, run it through DeepL, Google Translate, or ChatGPT, and hand in the result. The European Network for Academic Integrity defines translation plagiarism as "translations of work published in another language without acknowledgment"; researchers also call it cross-language plagiarism. Whatever the name, classical plagiarism detection is structurally blind to it.

Why plagiarism checkers miss translated text

Text-matching software compares character strings. Translation replaces every single one of them. In the most rigorous peer-reviewed test to date, nine researchers tested 15 detection systems across eight languages (Foltýnek et al. 2020). On translated documents, 14 of 15 tools scored between 0.0 and 0.5 out of 5. That's the worst result of any plagiarism type, with matches found "mainly in the references, not in the texts." Tellingly, the best performer was the only tool using semantic analysis.

Does Turnitin detect translated text?

Only partially. Its Translated Matching feature converts non-English submissions into English and then runs a standard comparison. That helps in one direction, for supported languages, but it remains weakest exactly where the risk is highest: carefully edited or human translations, and matches outside the English database.

Can AI detection close the gap?

No, and it can actively harm. AI detection answers a different question ("does this sound machine-written?"), not "where does this content come from?". Worse: when Weber-Wulff et al. (2023) tested 14 detectors, accuracy dropped by roughly 20 percentage points on machine-translated texts, and false positives rose sharply. With 6.9 million internationally mobile students worldwide, that is a recipe for wrongly accusing the very students writing in their second language. That is the opposite of academic integrity.

Semantic Source Search: the meaning survives translation

Here is the epiphany: translation destroys the words but preserves: the meaning, the claims, their sequence, the choice of sources. Modern multilingual language models map sentences into a shared semantic space, where a text and its translation sit close together even with zero matching strings. Mentafy's Semantic Source Search (S³) follows exactly this trail: it compares what a text says, not which characters it uses, so the connection to the original source remains visible across languages, just as it does for AI-paraphrased para-plagiarism.

And as always: a finding is not a verdict. A properly cited translation is legitimate scholarship. An uncited one is a question for the examiner — S³ makes sure the question gets asked.

Academic integrity in the AI era means using every signal

Cheating no longer leaves matching strings behind. An AI-ready integrity check therefore has to read every layer a text offers: semantics, sources, references, and where available the writing process itself. That is how academic writing stays assessable, and how higher education stays credible, in the age of AI.