Beating AI detectors by getting lost in translation on purpose

What it does

text-humanizer is a Python tool that attempts to make AI-generated text read as if a human wrote it. It runs prose through a multilingual pipeline—rewriting it with DeepSeek, translating it into Chinese, then Turkish, then optionally Japanese, and finally reconstructing it in the original language—to break the stylistic patterns that detectors like Turnitin and GPTZero look for. The project claims it preserves the original meaning while improving naturalness across several supported languages.

The interesting bit

Instead of simple synonym-swapping, the project weaponizes cross-lingual translation noise. Each hop through a structurally different language warps syntax and phrasing in ways that monolingual detectors struggle to trace, and a final DeepSeek pass smooths the accumulated artifacts back into coherent prose.

Key highlights

• Four-stage pipeline: DeepSeek rewrite, Google Translate to Turkish, optional DeepL hop from Turkish to Japanese, and DeepSeek final reconstruction back to the input language.
• Claims to bypass most AI detectors by distorting sentence structure through rule-based and neural translation differences.
• Supports English, Japanese, Chinese, Korean, German, French, and Spanish for input and output.
• Requires a DeepSeek API key and internet access; an optional DeepL key adds a second independent translation engine.
• MIT-licensed, though the core logic is essentially orchestration around external cloud APIs.

Verdict

Targeted at users who need to bypass AI-detection tools and are comfortable relying on external translation and LLM services. Anyone seeking a fully offline, self-hosted rewriter should keep scrolling.