Summary

Researchers introduce C-ReD, a comprehensive Chinese benchmark for detecting AI-generated text, built from real-world prompts rather than synthetic or homogeneous sources. Large language models (LLMs) can now produce highly fluent text, which brings convenience but also risks such as phishing and academic dishonesty, spurring work on detection algorithms and datasets. However, existing Chinese corpora for AI-text detection suffer from limited model diversity and data homogenization. C-ReD addresses these gaps by grounding generated text in authentic user prompts, enabling more realistic and varied training and evaluation data. Experiments show the benchmark supports reliable in-domain detection while also generalizing strongly to unseen LLMs and external Chinese datasets. The paper covers contributions from Chenxi Qing, Junxi Wu, Zheng Liu, Yixiang Qiu, Hongyao Yu, Bin Chen, Hao Wu, and Shu-Tao Xia, and is available on arXiv as 2604.11796.