Fundamental problems of validity in artificial intelligence: the role of generative artificial intelligence and its limitations

Abstract

This article examines fundamental issues of reliability in artificial intelligence systems, focusing on generative models widely used in scientific, educational, managerial, and social practice. The relevance of this study is determined by the expanding scope of generative AI and the increasing risks associated with the creation of substantively convincing, but factually incorrect, logically contradictory, or empirically unconfirmable results. The purpose of the study is to identify and theoretically understand the key factors determining the limitations of reliability in AI systems, as well as to determine the technological, epistemological, and applied barriers affecting the reliability, verifiability, and interpretability of the generated information. The study utilized methods of theoretical analysis of artificial intelligence models. The methodological basis was formed by theoretical-analytical and interdisciplinary approaches, which allowed us to consider the problem at the intersection of computer science, philosophy of science, epistemology, law, and ethics. The study includes scientific publications by Russian authors devoted to issues of trust in intelligent systems, the limitations of algorithmic models, algorithmic opacity, verification of results, and the specific features of generative artificial intelligence. The study found that the reliability of AI results is determined by a combination of factors: the quality and representativeness of training data, the opacity of neural network architecture, the ability of models to reproduce errors and biases, and the effect of the external plausibility of generated content, which can induce false trust in users. It is substantiated that generative AI does not understand content in the strict sense, but rather functions based on the statistical reproduction of patterns, which predetermines the possibility of systematic distortions. It is concluded that it is necessary to develop mechanisms for the verification, interpretation, and critical evaluation of AI results, especially in areas where decisions have significant social, legal, economic, and ethical consequences.