Laundering LLM prose through Chinese, Japanese, and Finnish

What it does

Humanize-Text is an open-source Python toolkit that tries to make AI-generated prose read as human. Its current “Standard Pipeline” takes English input, runs it through two DeepSeek rewrites—first into Chinese, then into Japanese—before translating it through Finnish via Google and Niutrans and finally reconstructing it in English. The stated goal is to break the statistical fingerprints that detectors like Turnitin and GPTZero look for while preserving the original facts.

The interesting bit

The core trick treats translation like a cryptographic mixer: each hop, especially the leap to Finnish across competing NMT engines, introduces compounding structural noise so that no single model’s signature survives the trip. It is essentially an elaborate game of telephone played across distant languages and APIs, with the authors claiming the result scores 9.1/10 on expert quality metrics.

Key highlights

• 4-step pipeline: DeepSeek rewrites (EN → Chinese → Japanese), then Google Translate (Japanese → Finnish) and Niutrans (Finnish → EN).
• Self-reported benchmarks: 100% key-information retention across 50 text pairs and an overall 9.1/10 expert quality score.
• 5 showcase examples with full intermediate traces; final outputs were labeled human by an unnamed AI detector, though confidence scores ranged from 0.7218 to 0.9997.
• Includes an n8n workflow and four reference methodologies under src/methodologies/ for research or customization.
• The open-source release covers only the “Standard” tier; “Advanced” and “Focus” tiers are reserved for the commercial Lynote.ai service.

Caveats

• The repo is essentially a funnel for the paid Lynote.ai platform, with the open-source version limited to the Standard tier.
• It requires live API keys for DeepSeek, Google Translate, and Niutrans, so it does not run offline or out-of-the-box.
• Detection-bypass claims and quality scores are self-reported; the evaluator and the specific detector model are unspecified in the README.

Verdict

Worth a look if you are building automation that needs to obfuscate AI-generated copy or researching adversarial text transformation. Skip it if you need a guaranteed, offline, self-contained solution, or if you find the constant upsell to Lynote.ai distracting.