Reliable Input Data as the Foundation for Trustworthy AI Applications

Abstract

we analyzed the fundamental limitations and risks of artificial intelligence (AI), with particular emphasis on large general-purpose models such as GPT. We showed that the era of extensive AI growth is approaching its limits because the supply of high-quality human-generated data is becoming exhausted. We considered AI-related challenges within the framework of ill-posed inverse problems and demonstrated that computational instability is an inherent mathematical property of such problems. We examined the theoretical foundations of the Transformer architecture and showed that its application to domains without formal syntactic structures, such as medicine and industry, lacks rigorous mathematical justification, raising concerns about the reliability of the resulting outputs. We analyzed examples from secondary and higher education and identified signs of stagnation associated with the delegation of intellectual tasks to generative AI systems. Based on our findings, we concluded that future AI development should shift from universal models toward specialized solutions supported by mathematical justification, verification of domain-specific source data, and mandatory human oversight in applications involving critical infrastructure.