larashero3-dotcom/lieflat-less-ai-tone

What it solves

This project addresses the "AI flavor" (AI-generated style) in Chinese text, which is often perceived as unnatural or formulaic. It aims to quantify what constitutes this style through linguistic research and provide a systematic way to remove these markers to make AI-generated content feel more human-like.

How it works

The project uses a corpus-based linguistic approach. It analyzed 2.8 million Chinese characters across 629 articles (300 AI-generated from five different models and 329 human-written). By testing 26 candidate features of "AI style," the project identified 11 statistically significant markers—such as specific punctuation patterns (excessive dashes), structural repetitions, and poor paragraph transitions (zero-anaphora comments at the start of paragraphs).

These findings are then converted into a set of rewrite rules (an Agent Skill) that can be applied to text. The system follows a "white-list" principle, only modifying the identified 11 markers while preserving the original meaning, facts, and overall structure of the text.

Who it’s for

• Content creators who want to remove the obvious markers of AI generation from their Chinese writing.
• AI agents and developers building writing assistants that need a more natural, human-like output in Chinese.
• Linguists interested in the quantitative analysis of LLM-generated text versus human writing.

Highlights

• Data-driven approach: Based on a statistical analysis of nearly 3 million characters to avoid reliance on subjective impressions.
• Model-agnostic: Analyzes multiple models (Claude, DeepSeek, Gemini, GPT, Kimi) to identify common and model-specific traits.
• Strict constraints: Ensures information conservation, meaning that no new facts are added and no original conclusions are deleted during the rewrite process.
• Executable rules: Provides a set of regex-based operators to identify and identify the "AI flavor" markers concretely.