Classifier Detects LLM-Generated Text Without Watermarks

On Jan. 10, 2026, Jin Zhu posts an arXiv preprint presenting a classifier to detect whether text is authored by large language models or humans. The detector, deployed on an online CPU-based platform, claims three novelties — no reliance on watermarks or model identity, improved discrimination, and enabling statistical inference — and reports higher accuracy, controlled type-I error, and computational efficiency.