Translation and AI Editing Failed to Fool Detectors: Doumate Raised the AI Share by 5–25% Afterward

Ask a student how to "beat the plagiarism check," and they will name two methods: ask a chatbot to edit the text with a special prompt, or run it through translation into English and back. The author of this experiment, published on Habr, decided to check whether this actually works. She also wanted to understand why detectors sometimes flag perfectly human work.

How the experiment was conducted

Five texts of 11–12 thousand characters each were used. The first is the author's 2021 grant application about wastewater treatment facilities. The second is its "twin," generated by Claude from the same technical brief. The other three are term papers from 2024 and 2025 and a 2026 graduation thesis (VKR) by one philology student. All the works were defended with a grade of "5" (the top mark).

Two techniques were applied to the texts. The first was AI editing by ChatGPT using a prompt: vary sentence length, change word order, avoid smooth transitions, and do not alter the meaning. The second was translation by Claude: RU→EN, EN→RU, and a double RU→EN→RU. The texts were checked with Doumate and GPTZero (model 4.1m). Antiplagiat was run only selectively: a full test was too expensive.

What the detectors showed

GPTZero and Antiplagiat rated the human text as human, 0% AI, after any processing. Doumate, however, returned 51.1% AI after double translation. The author says she found the cause and removed it, after which the score fell almost to zero. She promises details in a separate article, so for now this has to be taken on faith.

Things are duller for those looking to cheat with machine text. Doumate lowered the AI share by less than 3% after any rework. GPTZero shifted most from prompt-based editing: the probability of AI fell by 38%, almost by half. But the "generated" verdict remained.

A separate observation concerns language. The Russian-language Doumate confidently catches Russian generation, but rated the English generated text at 0% AI. GPTZero, by contrast, gives 100% on English and less precise percentages on Russian, although its verdict is correct. Each detector works better in its "native" language.

On the real term papers, the AI share according to GPTZero rose from 0% to 56% over three years, and according to Doumate from 12.5% to 70%. The academic supervisor doubted the 70% for the thesis. The author claims she managed to substantially lower that score without changing the content. "Reworking" in these papers more often raised the AI share in Doumate by 5–25%, and only in one case lowered it by 3%.

Limitations and practical takeaway

This is not a study with a representative sample. There are five texts, one author, the detector versions are not fully disclosed (Doumate's version is not specified), and tables with exact figures are not visible in the retelling. The mechanism behind Doumate's false positives is not explained and is postponed for the future. So the conclusions should be treated as a hypothesis, not as proven fact.

Three things are useful for those building or deploying such products. First, detectors react differently to edits: Doumate took a human text for machine-written after translation, while GPTZero was not fooled. Second, the language of the text affects reliability, and a detector for Russian cannot be applied to English-language works. Third, a single percent of AI share is no grounds for sanctions. If the score jumps after light processing, it needs to be rechecked with a second tool and a conversation with the author.

For students, the author's conclusion is straightforward: a machine will not make a text human. For universities and edtech teams, it is less obvious: instead of a percentage threshold, it is better to look at the share of "human" text and use several detectors. Until the causes of false positives are published, a Doumate check should be considered a guide, not a verdict.