Old N-Grams Never Die:Towards Identifying LLMs-Generated Text using Antique N-Grams

Usman Ahmed, Chaowei Zhang, Longwei Wang, Khan Hasanozzaman, Yi Zhu, Jipeng Qiang, Yizhe Wang, Yang Yi

Keywords: Antique N-Grams, LLM Paraphrasing, Machine-Text-Detection

Abstract: The proliferation of large language models (LLMs) has triggered an influx of AI-generated content, making robust detection of such content paramount for maintaining academic, journalistic, and regulatory integrity. However, the community has largely overlooked a time-tested resource that classical n-gram models, trained exclusively on human-authored corpora, may serve as a de facto gold standard for identifying machine-generated writing. In this paper, we build upon well-trained pre-AI N-Gram models to form the backbone of a lightweight AI-text detection system called \textbf{GramGuard}. Specifically, by generating paraphrased variants via temperature-controlled decoding from LLMs, we measure the shifts in log-likelihood, entropy, and token frequency variance between original texts and perturbed versions. These \emph{delta} features then feed into an ensemble classifier to yield interpretable decisions about authorship. Extensive experiments on PubMed, WritingPrompts, and XSum demonstrate that \textbf{GramGuard} matches or exceeds state-of-the-art detectors in performance and robustness. Our findings reaffirm the enduring value of pre-AI n-gram models and introduce a scalable, transparent solution for AI-text detection.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 9089