The article focuses on semantic and cognitive-pragmatic adequacy of neural machine translation in transmitting political media texts containing culturally marked invective units. The relevance of the study is determined by the widespread integration of neural machine translation systems into media production processes aimed at the rapid dissemination of information in multiple languages. The aim of the study is a comparative analysis of translations performed by professional translators of online media outlets and neural machine translations (Google Translate, DeepL, Yandex Translate) to identify typical machine translation errors in rendering culturally marked invective vocabulary and the need for their post-editing by humans. The research material includes fragments of political media texts in Russian, English, and Armenian containing implicitly and explicitly expressed culturally marked elements of verbal aggression (zoometaphor-invective, ethno-dysphemism, pseudo-ethnic nomination, pejorative vocabulary etc.), as well as their translations into the target languages. The source language varies depending on the specific case. A comparative-contrastive methodology was applied, combining discourse analysis, cross-cultural and translation-oriented approaches, to identify differences at the semantic and pragmatic levels, as well as at the level of culturally conditioned cognitive components of the text. It was revealed that neural machine translation systems produce systematic errors associated with the predominance of the literal translation principle, resulting in either the loss or, on the contrary, to the intensification of verbal aggression, the disruption of communicative impact, and the inadequate rendering of cognitive frames, dynamically shifting meanings, irony, sarcasm, and wordplay. The findings indicate a partial or complete absence of semantic, cultural and pragmatic adaptation in machine translation systems, which points to the limitations of these systems in conveying culturally and pragmatically marked textual content. The study demonstrates that a complete transition to neural machine translation systems without subsequent human post-editing is premature.