humanize-text is an open-source toolkit (lynote-ai/humanize-text on GitHub) that rewrites AI-generated English text to evade AI detectors such as GPTZero and Turnitin. Its standard pipeline chains four transformations: (1) a high-temperature (1.3) DeepSeek rewrite from English to Chinese, (2) a second DeepSeek pass from Chinese to Japanese with conversation history, (3) Google Translate from Japanese to Finnish, and (4) Niutrans back-translation from Finnish to English. By routing text across unrelated language families and multiple engines, the pipeline disrupts the statistical signals detectors rely on: low perplexity, low burstiness (uniform sentence lengths), and classifier fingerprints. The post explains how each detection metric is attacked, walks through the code structure and config, and documents practical pitfalls: Google Translate length limits, terminology drift, semantic drift (e.g., 'paradigm shift' degrading to 'attracted much attention'), and API cost/latency. It also reviews research on detection evasion (ESPERANTO, StyleShield, DetectGPT) and detector bias, including a Stanford finding that detectors misflag 61% of essays by non-native English writers. The author concludes the tool exposes the fragility of AI detection, raises academic-integrity concerns about its commercial positioning, and argues watermarking will eventually supersede statistical detection.
humanize-text: Washing AI Fingerprints Through a Four-Language Translation Chain
Summary
humanize-text is an open-source toolkit (lynote-ai/humanize-text on GitHub) that rewrites AI-generated English text to evade AI detectors such as GPTZero and Turnitin. Its standard pipeline chains four transformations: (1) a high-temperature (1.3) DeepSeek rewrite from English to Chinese, (2) a second DeepSeek pass from Chinese to Japanese with conversation history, (3) Google Translate from Japanese to Finnish, and (4) Niutrans back-translation from Finnish to English. By routing text across unrelated language families and multiple engines, the pipeline disrupts the statistical signals detectors rely on: low perplexity, low burstiness (uniform sentence lengths), and classifier fingerprints. The post explains how each detection metric is attacked, walks through the code structure and config, and documents practical pitfalls: Google Translate length limits, terminology drift, semantic drift (e.g., 'paradigm shift' degrading to 'attracted much attention'), and API cost/latency. It also reviews research on detection evasion (ESPERANTO, StyleShield, DetectGPT) and detector bias, including a Stanford finding that detectors misflag 61% of essays by non-native English writers. The author concludes the tool exposes the fragility of AI detection, raises academic-integrity concerns about its commercial positioning, and argues watermarking will eventually supersede statistical detection.
This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980697