Truth Mirror (2D Detection)

This code is for our paper "Decoupling Content and Expression: Two-Dimensional Detection of AI-Generated Text".

Paper | Demo

We are working on the demo and will update the link soon.

Brief Intro

We detect three levels of AI risk by decoupling a text into content and expression dimensions. The detector performs a binary classification within the two-dimensional space for each detection task. Our findings show that content is resistant to surface-level changes, which can serve as a key feature for detection.

Environment

• Python3.12
• Setup the environment: pip install -r requirements.txt

Experiments

Following folders are created for our experiments:

• ./exp_main -> experiments on English datasets from HART (main.sh).
• ./exp_langs -> experiments on CC News across five languages (langs.sh).
• ./exp_raid -> experiments on RAID (raid.sh).

(Notes: we share the data and results for convenient reproduction.)

Citation

If you find this work useful, you can cite it with the following BibTex entry:

@articles{bao2025decoupling,
  title={Decoupling Content and Expression: Two-Dimensional Detection of AI-Generated Text},
  author={Bao, Guangsheng and Rong, Lihua and Zhao, Yanbin and Zhou, Qiji and Zhang, Yue},
  booktitle={arxiv: xxx.xxx},
  year={2025}
}