A Systematic Analysis of Linguistic Features in AI-Generated Text Detection Across Domains and Models

Researchers conducted a large-scale empirical study analyzing 284 linguistic features across 27 LLMs and 10 text domains to identify which indicators reliably detect AI-generated text. The study found that while linguistic classifiers can distinguish AI from human text, most previously proposed indicators are context-dependent, with lexical richness measures proving the only robust signal across different models and domains.

This research addresses a critical gap in AI transparency by systematically evaluating which linguistic markers genuinely signal machine-generated content. The fragmented nature of prior findings—where different studies proposed different feature sets without consistent validation—created confusion for both researchers and practitioners trying to detect AI-generated text. This comprehensive analysis resolves that ambiguity by testing features across diverse contexts, providing empirical grounding for detection methods.

The finding that lexical richness remains robust while other proposed indicators prove context-dependent has significant implications for AI safety and content authentication. Current detection systems often rely on features that work in laboratory conditions but fail in real-world deployment across different model architectures and writing domains. This study essentially maps which linguistic signals have genuine predictive power versus which were artifacts of specific experimental conditions.

For platform developers, content moderators, and AI companies, the results suggest that simple lexical diversity metrics warrant investment as foundational detection components, even as more sophisticated models emerge. The cross-model and cross-domain validation methodology provides a replicable framework for evaluating future detection approaches. However, the research also implies that no purely linguistic approach may prove sufficient as LLMs continue evolving and improving their stylistic capabilities. Organizations relying on detection systems should recognize both the utility of robust linguistic signals and their limitations when deployed against increasingly sophisticated models.

• →Lexical richness is the only linguistic feature that reliably distinguishes AI-generated text across different models and domains.
• →Most previously proposed AI detection indicators are highly context-dependent and fail to generalize across different settings.
• →Classifiers based solely on linguistic features can effectively detect AI text, but their reliability varies significantly by domain and model.
• →The study validates 284 interpretable linguistic features across 27 LLMs, providing the most comprehensive cross-model analysis to date.
• →Current AI detection methods need validation across multiple domains and model families to ensure real-world effectiveness.