Toward Fine-Grained and Robust Human-AI Co-Authorship Detection

The researchers propose a discrete-category text provenance detection framework that grades human-AI co-authorship by degree of AI involvement and distinguishes surface-level polishing from rewriting at the discourse-structure level. The framework reveals how adversarial rewriting blurs the detector's decision boundary. Experiments across multiple datasets and generative models show that detectors trained only on benign data are substantially weakened by adversarial rewriting, whereas augmenting the training data with adversarial rewrites generated by state-of-the-art rewriting attack models yields consistent robustness gains and good cross-generator transferability.