New AI Detection Method Uses Alignment Imprints to Identify LLM-Generated Text

A new research paper introduces a method for detecting AI-generated text by analyzing the distributional imprints left during the alignment process of Large Language Models. The approach, called Log-likelihood Alignment Preference Discrepancy (LAPD), standardizes information-weighted statistics based on what researchers term the Alignment Imprint. This theoretical framework abstracts alignment—including fine-tuning and preference tuning—as a sequence of constrained optimization steps. The paper demonstrates that the log-likelihood ratio can decompose into implicit instructional biases and preference rewards. Existing likelihood-based detection methods often show unstable performance and sensitivity to content complexity. The research provides statistical guarantees that alignment-based statistics dominate traditional approaches, particularly in mitigating instability within high-entropy regions. Published on arXiv under identifier 2604.16923v1, this work addresses the challenging problem of AI text detection with a zero-shot methodology.