Розроблення підходу до запобігання виникненню інформаційного хаосу у чат-ботах з використанням породжувального штучного інтелекту

The object of this study is the linguistic constructions of queries to chatbots with Large Language Models (LLMs). The area of research is the emergence of information chaos during communication between the user and the chatbot, which leads to errors in forming a response to the query. It is assumed that the user and the chatbot are separate complex systems, the events and actions of which are difficult to predict for a long period. Behavior models of complex systems are subject to the influence of chaos theory. To demonstrate this, the work used one of the simple mathematical problems with a logical component. The Copilot and ChatGPT 4o mini chatbots that were studied gave erroneous results in response to a query for the task. The error occurred when generating a query due to the introduction of a logical component. A similar process was represented by a system of differential equations solving which establishes clear rules for obtaining an accurate answer to the query. To submit a request from the user, an approach has been proposed that makes it possible to break down the information block of the request by constructing piecewise linear attractors. That is, paired semantic expressions are formed with the formation of a request cleared of information noise. The problem is solved by presenting a methodology for controlling the selection of substitute words, based on the operations of generating the next substitution and calculating the number of the given substitution. According to the devised approach, the best options for a request to the Copilot chatbot were obtained in 182 characters or 48 words, numbers, and special characters. For the ChatGPT 4o mini chatbot, such a request consisted of 219 characters. The proposal could be used in practical activities to improve chatbot technologies and form key data sets in artificial intelligence systems, which would further make it possible to avoid errors when solving problems with a logical component