Text-Humanizer sends AI drafts through Turkish and back to strip the AI patterns, 722 stars in three days

Any writer who has pasted a chatbot draft into an AI-detection tool like Originality or GPTZero has watched the score come back saying the text is machine-written. A new open-source project on GitHub tries to solve that problem with translation.

Text-Humanizer, released on 23 September 2026 by an author using the handle asokurasu, is an MIT-licensed Python project that rewrites AI-generated text by running it through a chain of translation and paraphrase steps. As of 26 September 2026 the repository carries 722 stars and 83 forks, all gained in three days, and the code has been forked into at least a dozen personal repositories in the same period.

What the tool does

The pipeline has four steps, all named on the project's README. First, the input text is passed to DeepSeek's language model with an instruction to rewrite it and, in the same pass, translate the result into Chinese as an intermediate form. Second, that Chinese output is sent through Google Translate into Turkish. Third, and only if the user has a DeepL API key, the Turkish text is translated again into Japanese. Fourth, DeepSeek translates the final version back into the original input language and cleans up the artefacts that always accumulate when a passage is bounced between languages.

The author says the point is what happens on the way. Each translation step rearranges word order, splits or fuses clauses, and picks different synonyms, so by the time the text reaches the last stage it no longer carries the sentence rhythm the source model produced. The README table lists seven language codes for input and output: English, Japanese, Chinese, Korean, German, French, and Spanish. A summary line at the top of the README says eight are supported, so the exact number is unclear from the source page.

What the author claims and does not

The README lists five features. Three are practical: the pipeline preserves the original meaning, it is adaptable to different writing styles and tones through a temperature setting, and it is free and open source. The other two are what will draw traffic. The author says the tool "bypasses most of AI-detectors" and that DeepSeek's temperature set to 1.3 gives the best results. Neither claim is measured on the page and no side-by-side scores are shown. Both are the author's own account.

The repository is a single Python entry point plus a TOML config file, licensed MIT, and requires DeepSeek and Google Translate keys to function. DeepL is optional. There is no hosted version, no cost from the project itself, and no examples of before-and-after text checked into the repo.

Why the timing matters

Search engines now weight machine-authored content lower, and AI answer engines use the same classifier signals to decide what they quote. A market of paid humanising services has grown up alongside them, charging by the word. Text-Humanizer is a free, self-hosted answer to that same problem. The cost moves from a per-word subscription to two API keys the writer already has, and the code is auditable, which matters for teams whose legal or brand rules ban sending drafts through a third-party rewriter.

The catch is quality control. A four-step translation round trip preserves meaning most of the time. It also introduces small factual changes: a proper noun spelled two ways, a metric unit converted and not reverted, a specific number rewritten as a range. A team using this tool has to read the output word by word before publishing, or the humanisation costs more in editing than it saved in subscription fees.

Source

• asokurasu/text-humanizer on GitHub
• GitHub REST API metadata for the repository (star count, licence, creation date, verified 2026-09-26)