korcarc/text-humanizer

The core idea here is that AI detectors are pattern-matchers, and the best way to defeat a pattern-matcher is to route text through enough unrelated linguistic systems that the original statistical fingerprint dissolves. text-humanizer does exactly that: it chains DeepSeek, Google Translate, an optional DeepL pass, and DeepSeek again, bouncing content through Chinese, Turkish, and optionally Japanese before reconstructing the final output in the original language.

The mechanism is worth understanding on its own terms. Each translation hop is chosen deliberately. Chinese as an intermediate step breaks English syntactic habits at the clause level. Turkish, a highly agglutinative language with SOV word order, introduces structural distortion that neither English nor Chinese would produce. Japanese via DeepL adds a second independent translation engine with its own statistical biases. By the time DeepSeek reconstructs the target language, the sentence structures, phrasing rhythms, and information sequencing have been scrambled enough that detector models trained on typical LLM output struggle to recognize the source.

The README is candid about what this is for: bypassing AI detectors. There is no pretense that the goal is purely stylistic improvement, though meaning preservation is a real design constraint — the final DeepSeek pass is specifically tasked with removing accumulated translation artifacts and restoring readability.

Practically, the setup requires a DeepSeek API key; the DeepL step is optional and, per the README, improves linguistic diversity by introducing a second independent translation engine. Eight languages are supported for input and output. Temperature is configurable — the docs recommend 1.3, higher than typical generation defaults, which makes sense if the goal is stylistic variance rather than factual precision.

The fragility here is obvious: AI detectors are not static. The specific translation chain that defeats today's detectors may not defeat next year's, especially as detector developers start training on outputs from exactly these kinds of multi-hop obfuscation approaches. The repo makes no claims about long-term robustness.

What is genuinely interesting from a technical standpoint is the use of typologically distant languages as noise injectors rather than just round-tripping through similar languages. That reflects real awareness of how syntactic transfer works across language families. Whether it translates to durable evasion is a separate question from whether the architecture is clever — and it is.