The paper deals with Large Language Models (LLM). We present a historical overview of the development of text generation algorithms. The aim of the paper is to show the main properties and limitations of services based on Large Language Models when their results are used in scientific and educational texts. We touch on concepts such as sampling from distribution, Recurrent Neural Network, Attention mechanism, Transformer architecture and provide guidelines for the ethical use of generated texts in scientific and academic papers. We have presented the material in a popular form, so that it is possible to understand the principles of generative services with general erudition and certain computer skills. At the same time, the key concepts are provided with references that will allow readers to delve into the area on their own if necessary.