Title: A Comprehensive Study on the Generalizability of Linguistic Markers in AI Text Detection Across Various Models and Fields

Abstract: Interpretable linguistic characteristics present a viable method for clarifying the origins of machine-generated text, a capability that is particularly valuable for non-specialist audiences. Nevertheless, current research regarding which specific features consistently signal Large Language Model (LLM) output is scattered across disparate feature sets, models, and textual contexts. To bridge this knowledge gap, we performed a large-scale empirical investigation to evaluate the stability of linguistic cues used to identify AI-generated content. This study examined 284 distinct interpretable linguistic features derived from the outputs of 27 different LLMs, spanning ten varied text domains, while assessing performance under both cross-model and cross-domain generalization conditions. Our findings confirm that classification models relying exclusively on linguistic features are capable of accurately differentiating between human-written and AI-produced text. However, we observed that numerous previously suggested indicators are highly sensitive to context, with the notable exception of lexical richness metrics, which demonstrated consistent robustness across different model architectures and text categories. These outcomes highlight which linguistic signals maintain their efficacy across diverse contexts, thereby establishing a basis for more dependable and transparent analysis of AI-generated language.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC