Ten days after ChatGPT, they started building detectors

What it does This repository hosts HC3, a bilingual English-Chinese corpus that places human answers side-by-side with ChatGPT answers across medicine, finance, law, and open QA. It also releases three fine-tuned RoBERTa detectors—single-text, QA-pair, and linguistic-feature variants—as interactive demos and downloadable model weights.

The interesting bit The team started the project ten days after ChatGPT launched, racing to build open detection tools while most of the world was still writing its first prompts. They attacked the problem from two sides at once: standard neural classifiers and a separate linguistic-feature model.

Key highlights

• HC3 covers English and Chinese across domains including medicine, finance, and law
• Three detector variants: isolated text, question-answer pairs, and linguistic-feature analysis
• Model weights and datasets are available on Hugging Face and ModelScope
• Built on standard roberta-base and chinese-roberta-wwm-ext backbones

Caveats

• Several source datasets in the corpus carry unknown or unverified licenses; the authors note they are actively seeking clarification from original publishers

Verdict Worth a look if you are researching AI-generated text detection or need a bilingual human-vs-LLM corpus. Look elsewhere if you need a polished, production-grade API with guaranteed uptime.