This project addresses the growing challenge of academic papers being falsely flagged by AIGC detection tools. It offers a Claude Code skill designed to reduce the detection rates of AI-generated content in academic writing by making text exhibit statistical features of human authorship. The target audience includes researchers and students seeking to mitigate the risk of their papers being misidentified as AI-generated, thereby preserving the integrity of their work.

How It Works

The core approach integrates technical principles from major AIGC detectors (like 知网, 万方, PaperPass) with AI writing detection methodologies derived from sources such as Wikipedia's "Signs of AI writing." It employs a novel three-round reduction protocol: first, it subtracts AI traces through word-level replacement, sentence restructuring, and paragraph adjustments to ensure a modification rate exceeding 40%. Second, it injects human-like features by varying sentence length, introducing uncertainty, and adding operational details. Finally, it performs an anti-AI audit, scanning for 10 specific deep AI trace patterns and utilizing an automated script (aigc_scan.py) to check seven distinct AI feature dimensions correlated with detector algorithms. This multi-stage process aims to reduce false positives for human-written text rather than simply masking AI output.

Quick Start & Requirements

Installation involves cloning the repository: git clone https://github.com/xiaofenggan01/aigc-reduce.git ~/.claude/skills/aigc-reduce The skill is automatically triggered within the Claude Code environment when specific keywords related to AI detection reduction are used. The aigc_scan.py script requires Python 3. No specific hardware (GPU/CUDA) or advanced Python versions are mentioned for the core skill functionality.

Highlighted Details

• Targets fundamental limitations of AIGC detectors, including high false positive rates for quality writing and cross-platform result inconsistencies.
• Employs a three-round protocol: AI trace subtraction, human feature injection, and self-auditing.
• Prioritizes a modification rate greater than 40%, deemed essential for effective evasion, contrasting with less effective light modifications.
• Features an automated scanning script (aigc_scan.py) that analyzes text across seven dimensions linked to specific detector mechanisms.
• Methodology synthesizes empirical research from academic reports, Wikipedia, and community-driven prompt engineering techniques.

Maintenance & Community

No specific details regarding maintainers, community channels (like Discord/Slack), or ongoing development efforts are provided in the README.

Licensing & Compatibility

The project is released under the MIT license. This permissive license generally allows for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

The tool is explicitly positioned as a means to reduce false positives for originally human-written papers, not as a tool for academic dishonesty. The README warns that full AI rewriting can paradoxically increase detection rates. Effectiveness is contingent on achieving the >40% modification threshold. The skill's integration with "Claude Code" suggests potential limitations for standalone use outside that specific platform.