Alignment Imprint: Zero-shot AI-generated text detection based on provable preference differences

The researchers propose a zero-shot AI-generated text detection method: the alignment process of LLMs (including fine-tuning and preference tuning) leaves a measurable distributional imprint, called the alignment imprint. They theoretically prove that alignment-based statistics outperform Fast-DetectGPT; the standardized information-weighted statistic LAPD built on this imprint improves by 45.82% relative to the strongest existing baseline, achieving large and consistent gains across all settings.