This study examines the inherent limitations of Artificial Intelligence Translation (AIT) through the lens of universal grammar and context-oriented translation theories, with particular reference to Wittgensteinian philosophy of language. Through analysis of translation errors in Russian-Korean AIT outputs, this study argues that certain limitations stem from fundamental algorithmic constraints rather than merely technical implementation challenges. While AIT systems effectively leverage vector databases to store, manage, and optimize word relationships in context, allowing for increasingly fluid translations, their operation remains fundamentally aligned with Wittgenstein's concept of language games. However, significant translation errors persist in pragmatic dimensions, including narrator intention, idiomatic expressions, nuance, and rhetorical connotations. In essence, AIT systems struggle to capture what Benveniste terms "language outside the text" - the domain of enunciation. Despite their considerable capabilities, AIT systems remain confined to text-to-text operations, unable to fully engage with the broader space of communication between narrators and auditors where idioms and significations are dynamically generated. While AIT can translate the narrative, it seems unable to grasp narration itself; it can translate utterances but not the act of enunciation - a capability that remains uniquely human.