Luminol-AIDetect: Zero-Shot Detection of Machine-Generated Text via Perplexity Shifts
Researchers propose Luminol-AIDetect, a zero-shot statistical method for detecting machine-generated text (MGT) by exploiting structural fragility in autoregressive language models. The approach applies randomized text shuffling and measures perplexity shifts, which are more dispersed for MGT than for human-written text. This model-agnostic technique avoids reliance on specific generation fingerprints and requires only a few scalar features for classification. The method is detailed in arXiv preprint 2604.25860.
Key facts
- Luminol-AIDetect is a zero-shot statistical approach for MGT detection.
- It uses randomized text shuffling to disrupt coherence.
- Perplexity-under-shuffling dispersion differs between MGT and human text.
- The method is model-agnostic and does not rely on specific generation fingerprints.
- It requires only a handful of perplexity-based scalar features.
- The research is published on arXiv with ID 2604.25860.
- The hypothesis is that autoregressive LLMs have structural fragility.
- The approach exposes fragility through coherence disruption.
Entities
Institutions
- arXiv