Does translating AI text into another language and back help?
A little, but at a price that is often not worth paying. A round trip through another language breaks up the original sentence structure, so some of the machine rhythm disappears. At the same time it simplifies the text, averages out the vocabulary and can shift the meaning and the technical terms — and detectors read averaged-out text much like machine text. The deep rewrite in DetekceGPT uses this step only as one of several and discards the result if the length of the text changes by more than a fifth.
Why people try it
A model writes in set patterns and translation disrupts them: sentences are put together differently, words are swapped and favourite turns of phrase that have no exact counterpart in the target language disappear. After translating back, the text looks less like output from a chat at first glance.
It is quick and free, which is why this approach is often recommended on social media. Less is said about what gets lost along the way.
What you lose
Translation averages a text out. Sentences settle into a similar length, the vocabulary gets simpler and idioms disappear — and that is exactly what detectors read as signs of machine writing. The result can therefore come out much like the original, just for a different reason.
The second risk is meaning. A term gets translated with a synonym that means something else in the field, set phrases fall apart and a subtle difference in a claim is lost. In an academic text that is a more serious problem than a mechanical style.
How the deep rewrite does it
The deep rewrite in DetekceGPT uses translation as one step. The text is first rewritten according to the text type, then goes through another language and back — an English text through German, Czech, Slovak and German texts through English — and the result is used only if the word count does not change by more than a fifth. Otherwise the translation is discarded and the standard rewrite carries on.
After the translation, small human irregularities are added to the text and the result goes through checks of rhythm, phrases, repetition and length. Even so, the deep rewrite holds technical terms less well, which is why it is optional and best left switched off for an academic text.
FAQ
Does translating through several languages help?
Each further translation averages the text out more and raises the risk that the meaning changes. It will not reliably remove the machine rhythm.
Is a translated text still a text from AI?
Yes. Translation changes the language and the structure, not the origin. If you are supposed to declare your use of AI, translation changes nothing about that.
Can I use the deep rewrite without an account?
No, it is not in the demo without signing in. You switch it on in humanisation with a free account.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection