Detecting AI-generated text requires distinguishing not just human vs. AI, but also separating what was written from how it was written—content and expression can have different origins.
This paper tackles detecting mixed-origin text—where content and expression come from different sources (e.g., AI-written ideas expressed by humans). Instead of binary human-vs-AI classification, the authors propose D2C-Routing, which separately identifies content origin and expression origin, then combines them to classify four collaboration types.