Анотація: (English) Hallucinations in large language models (LLMs) are a systemic problem that manifests itself when models generate information that does not correspond to the ground truth or input data. This phenomenon significantly limits the application of LLMs in mission-critical domains such as medicine, law, research, and journalism, where the accuracy and reliability of information are of utmost importance. This paper provides a comprehensive analysis of three key factors that contribute to hallucinations: issues related to the quality and structure of training data; architectural features of transformer models that predispose them to error accumulation; and the lack of built-in fact-checking mechanisms, due to which models rely solely on statistical regularities. Each of these factors is discussed in detail using relevant research, and potential solutions are proposed. The paper includes three dedicated graphs that visualize the relationship between various model parameters and the occurrence of hallucinations. The results of the study indicate the need for a comprehensive approach to improving LLM, including both improving data preprocessing methods and modifying the model architecture and introducing additional verification mechanisms.