Humanize translated text? Why machine translation gets flagged as AI

Because a translation engine is a language model, and its output has the regularity detectors look for. Weber-Wulff and colleagues built a test set of 54 documents for 14 detection tools in 2023, and one category, which they called 02-MT, was human writing in Bosnian, Czech, German, Latvian, Slovak, Spanish and Swedish, about 10,000 characters each, translated into English with DeepL in 3 cases and Google Translate in 6 (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). None of the source texts had ever been online.

The detectors classified the hand-written English documents correctly 96 percent of the time. On the translated documents, accuracy "dropped by 20%", and the authors wrote that "apparently, machine translation leaves some traces of AI in the output, even if the original was purely human-written". Six of the fourteen tools produced false positives at all, and the paper says the risk rose "dramatically" for machine-translated texts; for GPTZero, half of its positive classifications on human and translated documents would have been false accusations.

The authors drew the conclusion that matters here in their discussion: using machine translation "can lead to a higher number of false positives, leaving L2 students (and researchers) at risk of being falsely accused of unauthorised content generation when using machine translation to translate their own texts". A translated text sits closer to the AI class than to the human one, and the writer did nothing but translate.

Yes, and the mechanism is the same one. Detectors score perplexity, how predictable each word is given the words before it. Liang and colleagues ran seven detectors over 91 human-written TOEFL essays in 2023 and found an average false positive rate of 61.22 percent, with the essays every detector flagged showing significantly lower perplexity than the rest (Liang et al. 2023, Patterns, read 9 November 2026). A second-language writer uses a smaller, more expected vocabulary; so does a translation engine choosing the most probable English rendering of each phrase.

The same paper shows what moves the reading. When the authors prompted a chatbot to enrich the vocabulary of the TOEFL essays, the false positive rate fell from 61.22 to 11.77 percent. That is the operation a humanizer performs, and it is also the operation that changes what a text says, because the enriched word is not always the one the writer meant. The non-native English page covers the writer's side of this; translation is the same problem with a machine in the middle.

Not if the text is yours, and the reason is evidence. A translated document that gets flagged has a defence a detector cannot see: the original-language file, its edit history, the translation tool's name and the date. A humanizer run over the translation removes that chain, because the submitted text no longer matches the translation, and it adds the risk every rewrite carries, a changed qualifier, a swapped term, a dropped clause. The rubric this site uses lists twelve of those faults, and the meaning score it feeds carries 35 of 100 points on the method page.

Weber-Wulff and colleagues' own conclusion was procedural rather than technical: detection reports "cannot be used as the only basis for reporting students", and a student accused on a detector score alone "would have no possibility for a defence" because the tools provide no evidence. The writer who keeps the source text has the defence the paper says is missing. The false positives guide sets out what to ask for and what to show.

Three vendors state a list and most state none. Grammarly's AI Humanizer page says the tool works in "English, Spanish, French, German, Portuguese, and Italian" (Grammarly, AI Humanizer, read 9 November 2026). HIX Bypass's pricing page lists "50+ Languages Supported" as a feature of every plan without naming them (HIX Bypass, pricing, read 9 November 2026). Netus AI's pages state 36 languages across its Bypasser, Paraphraser and Summarizer, read 9 November 2026. Clever AI Humanizer, Phrasly, StealthGPT, Undetectable AI, Walter Writes AI, WriteHuman, GPTHuman, QuillBot and BypassGPT state no language list on the pages checked on 8 and 9 November 2026.

A language list says the tool accepts the input. It says nothing about what the rewrite does to a text in that language, and no vendor publishes a per-language result. The languages page lists the tools by stated language support for readers who write in something other than English and want a rewrite in that language rather than a translation.

The published method's first run uses an English-only input set, so the translation question is answered on this page from the two studies rather than from a run; the method's six human-written control texts report each detector's false positive rate on the same genres, which is the figure a translated writer needs when a flag arrives. No detector score should be the only evidence in a misconduct case, and a translated text is the case where that line matters most. The rest of this series is in the guides hub.