2026 年 17 巻 3 号 p. 912-931
The rapid evolution of Large Language Models (LLMs) has necessitated robust technologies to distinguish AI-generated text from human-written content. However, existing detection methods often rely on semantic content words, causing significant performance degradation when applied to Out-of-Distribution (OOD) domains where the topic or writing style differs from the training data. To address this limitation, this paper proposes a domain-agnostic detection framework that focuses on the structural degeneration observed in Function Words (FW). We hypothesize that while content words vary by topic, the probabilistic maximization bias of AI models manifests universally as excessive typicality in function word usage. We introduce two approaches using the RoBERTa encoder: FW-RoBERTa, which utilizes averaged hidden state vectors of function words, and FW-PLL, which utilizes function-word-restricted Pseudo-Log-Likelihood scores. By filtering out content words, our method isolates the syntactic structure from explicit semantic content. Extensive experiments across multiple domains (arXiv, XSum, WritingPrompt) and generative models, ranging from GPT-3.5 Turbo to the state-of-the-art GPT-5.1 and Gemini 3 Pro, demonstrate that our approach maintains high accuracy in OOD environments where conventional baselines fail. These results confirm that context-aware function word features capture the intrinsic, invariant differences between human and AI text.
