Persuasive and misleading language has long been a powerful tool in shaping public opinion, guiding consumer behavior, and influencing political discourse. The complexity of detecting subtle rhetorical strategies, particularly at the token level, presents a significant challenge for traditional methods of text analysis. The novel approach developed in this study leverages token-level processing within transformer-based models to classify words based on their persuasive and misleading potential, providing a granular perspective on language manipulation. Through comprehensive experiments, the analysis demonstrated that tokens linked to sentiment polarity, lexical complexity, and positional importance play key roles in shaping the rhetorical impact of texts. This method provides an efficient and scalable solution for automated content moderation, political discourse analysis, and advertising regulation, with applications extending to media analysis and misinformation detection. The integration of attention mechanisms and contextual embeddings offers a detailed view into how language functions at a deeper structural level, positioning this framework as a significant advancement in automated text analysis.