Non-Parametric Machine Text Detection via Multi-View Gaussian Processes

The article presents a non-parametric machine text detection framework utilizing multi-view Gaussian processes, which aggregates diverse feature signals from documents to enhance detection resilience against adversarial attacks like paraphrasing and style transfer. This approach allows for improved performance by requiring adversaries to simultaneously overcome multiple detection axes, while also providing calibrated probabilities and abstention capabilities for out-of-distribution inputs. Evaluations on benchmarks such as DetectRL, RAID, and PAN2025 show that this method outperforms existing models, making it a significant advancement for practitioners focused on robust AI detection systems.