Alignment Imprint: Zero-Shot AI-Generated Text Detection via Provable Preference Discrepancy

2026-04-18

AI Summary

Current likelihood-based methods for detecting AI-generated text exhibit sensitivity to content complexity and suffer from unstable performance. This work theoretically demonstrates, for the first time, that the alignment process of large language models leaves measurable distributional imprints, and leverages this insight to propose LAPD—a zero-shot detection method based on alignment-preference information-weighted statistics. By modeling alignment-induced distributional shifts, formulating a constrained optimization abstraction, and decomposing the log-likelihood ratio, LAPD constructs a normalized detection score with statistical guarantees. Extensive experiments show that LAPD consistently outperforms the strongest baselines across diverse settings, achieving a relative performance gain of up to 45.82% and significantly enhancing both accuracy and robustness in AI-generated text detection.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Detecting AI-generated text is an important but challenging problem. Existing likelihood-based detection methods are often sensitive to content complexity and may exhibit unstable performance. In this paper, our key insight is that modern Large Language Models (LLMs) undergo alignment (including fine-tuning and preference tuning), leaving a measurable distributional imprint. We theoretically derive this imprint by abstracting the alignment process as a sequence of constrained optimization steps, showing that the log-likelihood ratio can naturally decompose into implicit instructional biases and preference rewards. We refer to this quantity as the Alignment Imprint. Furthermore, to mitigate the instability in high-entropy regions, we introduce Log-likelihood Alignment Preference Discrepancy (LAPD), a standardized information-weighted statistic based on alignment imprint. We provide statistical guarantee that alignment-based statistics dominate Fast-DetectGPT in performance. We also theoretically show that LAPD strictly improves the unweighted alignment scores when the aligned and base models are close in distribution. Extensive experiments show that LAPD achieves an improvement 45.82% relative to the strongest existing baselines, yielding large and consistent gains across all settings.

Problem

• AI-generated text detection
• zero-shot
• alignment
• likelihood-based detection
• distributional imprint

Innovation

• Alignment Imprint
• Zero-Shot Detection
• LAPD
• Preference Tuning
• LLM Alignment