<p>The rapid increase in blogs and opinionated online content has led to a growing focus on automated detection of political article orientation in fields like academia and national security. Political discourse, especially in Arabic texts, contains many ambiguous or indeterminate phrases, expressions with dual meanings, and nonstandard forms of language that lead to uncertain or conflicting interpretations. This challenge can be seen as a subset of text classification, where machine learning models must not only be sensitive to hyperparameter settings but also effectively capture the complexities and uncertainties of natural language semantics, addressing ambiguity, vagueness, and indeterminate expressions in the text. This paper introduces an intelligent system for detecting the political orientation of Arabic articles using an uncertainty-aware CatBoost classifier, enhanced by multilevel feature extraction to improve discrimination between categories. The system utilizes a neutrosophic loss function for CatBoost, designed to address the uncertainties and indeterminacies inherent in political sentiment detection. By utilizing neutrosophic logic, the approach extends traditional binary logic to explicitly handle uncertainty and indeterminacy in political texts. By employing a dataset of political Arabic texts from diverse sources—categorized into conservative, reform, and revolutionary opinions—the study shows that the neutrosophic-based CatBoost classifier achieves an impressive accuracy of 98.14%, outperforming other common text classification models.</p>