Title: Luminol-AIDetect: Rapid Zero-Shot Detection of Machine-Generated Text via Perplexity Shifts Induced by Text Shuffling

Abstract: Effective detection of machine-generated text (MGT) necessitates the identification of structural invariants common to various generation models, moving away from reliance on model-specific signatures. We posit that although large language models maintain strong local semantic coherence, their autoregressive architecture introduces a distinct structural vulnerability when compared to human composition. To exploit this, we introduce Luminol-AIDetect, an innovative zero-shot statistical framework that reveals such fragility through coherence disruption. Our method employs a straightforward randomized shuffling technique, showing that the consequent change in perplexity acts as a robust, model-independent discriminator. Specifically, MGT exhibits a unique dispersion pattern in perplexity under shuffling, which contrasts sharply with the more stable structural variations found in human-authored content. Luminol-AIDetect utilizes this differentiation to drive its classification logic: it extracts a small set of scalar perplexity features from both the original input and its shuffled counterpart, subsequently performing detection through density estimation and an ensemble-based prediction strategy. In evaluations spanning 18 languages, 11 types of adversarial attacks, and 8 content domains, Luminol-AIDetect achieves state-of-the-art results, delivering up to a 17-fold reduction in false positive rates (FPR) at a lower computational cost than existing methods.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC