Sending AI text on a multilingual round trip to beat detectors

What it does

text-humanizer is a Python orchestration script that feeds AI-generated text through a multilingual rewriting pipeline to evade detection tools like Turnitin and GPTZero. It supports eight languages and aims to preserve meaning while altering sentence structure and stylistic patterns. The heavy lifting is done by external services, not local code.

The interesting bit

The actual mechanism is a deliberate game of telephone: DeepSeek rewrites the text and translates it to Chinese for structural variation, then Google Translate sends it to Turkish, an optional DeepL leg pushes it into Japanese, and DeepSeek finally reconstructs the original language. The theory is that this cross-lingual distortion scrubs the stylistic fingerprints that detectors hunt for.

Key highlights

• Four-stage pipeline using DeepSeek, Google Translate, and optional DeepL to rewrite text.
• Claims to break AI detectors by introducing cross-lingual syntactic distortion.
• Supports eight languages: English, Japanese, Chinese, Korean, German, French, and Spanish.
• Requires a DeepSeek API key; DeepL is optional.
• MIT licensed.

Caveats

• Completely dependent on external cloud APIs, so it cannot run offline.
• The “bypasses most AI detectors” claim is presented without benchmarks or independent verification.
• The README’s description of the pipeline contains a slight logical gap: Step 1 outputs Chinese, yet Step 2 is labeled “EN → TR,” leaving the exact intermediate flow ambiguous.

Verdict

Researchers probing AI detector weaknesses will find the pipeline concept interesting; everyone else should note it is an API-dependent script explicitly built for evasion, not general-purpose editing.