Luminol-AIDetect: Zero-Shot Machine-Generated Text Detection via Perplexity Under Text Shuffling

arXiv·medium signal

Proposes a model-agnostic, zero-shot approach to detecting machine-generated text by measuring how perplexity changes when text is randomly shuffled. The key insight: LLMs produce text with strong local coherence that is structurally invariant to the specific generating model, and this local coherence signature breaks distinctly under shuffling compared to human-written text. No model-specific fingerprints needed, works across generation models, and computationally lightweight — a practical detector for content moderation at scale.