The text examines the evolution of artificial intelligence (AI) technologies in the field of translation, highlighting their role in expanding capabilities for processing scientific-technical and literary texts. Despite advancements in machine translation (e.g., seq2seq and transformer models), AI's mathematical methods fundamentally differ from human approaches, leading to distortions in semantics, cultural and emotional nuances, particularly in literary texts and simultaneous translation. Limitations of AI are noted in recognizing context, emotional intensity, and culture-specific features, including idioms and metaphors. Special attention is paid to ethical issues: bias, context manipulation (e.g., Wikipedia examples), and the need to adapt models to cultural, legal, and religious databases. As a solution, the modifying the attention mechanism in transformer architectures to focus on integrating lexicographic resources and interdisciplinary data to enhance translation accuracy and ethical correctness proposed.