An Ethical and Rhetorical Framework to Analyze LLM-generated Texts

The rapid diffusion of Large Language Models (LLMs) is transforming the production and circulation of textual content, raising important linguistic, cognitive, and ethical questions. As AI-generated texts increasingly populate digital environments, understanding their rhetorical and stylistic characteristics becomes essential to assess their influence on readers and on the broader information ecosystem. This paper proposes an ethical and rhetorical framework for the analysis of LLM-generated texts based on a reader-oriented approach. The framework identifies a set of rhetorical functions (trustworthiness, apologetic, proximity, diversification, ambiguity, emphasis, explanatory, poetic, fairness, structure) and links them to specific rhetorical figures, providing a structured method to evaluate how linguistic strategies shape readers’ perceptions and responses. Methodologically, the study integrates theoretical modeling with qualitative close reading, combining deductive framework construction with inductive textual analysis. The framework is tested through the analysis of two corpora of AI-generated texts, consisting of argumentative essays and narrative short stories. The results highlight recurrent rhetorical patterns, including the overrepresentation of emphatic and structural functions in argumentative texts and the prevalence of figurative language with limited semantic depth in narrative outputs. These findings suggest that LLMs tend to reproduce recognizable rhetorical patterns while often relying on formal persuasion strategies rather than conceptual complexity. The proposed framework is designed as an open, scalable, and integrable model that can be progressively refined and expanded through further empirical applications across different textual typologies and communicative contexts.