C-ReD: A Comprehensive Chinese AI-Generated Text Detection Benchmark Based on Real Prompts

The researchers propose C-ReD, a comprehensive Chinese AI-generated text detection benchmark built on real prompts, to address risks such as phishing and academic misconduct brought by content generated by large language models. In response to the challenges of limited model diversity and homogeneous data in existing Chinese corpora, the paper's experiments show that this benchmark not only supports reliable in-domain detection but also generalizes to unseen LLMs and external Chinese datasets, filling the gaps of previous Chinese detection benchmarks in model diversity, domain coverage and prompt authenticity. The related resources are available on GitHub (https://github.com/HeraldofLight/C-ReD）开源。