New Multi-View Gaussian Process Detects Machine-Generated Text Robustly.

Existing machine text detectors often fail under adversarial conditions such as paraphrasing or targeted style transfer, as these attacks can suppress specific detection signals. However, a document typically contains multiple complementary signals, including stylistic, likelihood, rank-order, and structural features, which an attack might not simultaneously defeat. To address this vulnerability, researchers have developed a multi-view, non-parametric detection framework. This system extracts diverse feature views from a single document and then aggregates the evidence from each view using a Gaussian process ensemble. This multi-faceted approach significantly raises the bar for adversaries, requiring them to overcome multiple independent detection axes simultaneously. A key advantage of the Gaussian process formulation is its ability to provide calibrated probabilities and principled abstention for out-of-distribution inputs. This feature is crucial for reliable deployment in high-stakes environments where confidently incorrect predictions are unacceptable. Evaluations on benchmarks like DetectRL, RAID, and PAN2025 demonstrate that this multi-view detector maintains strong performance against various attacks, outperforming current methods.