Epistemic Limitations of Large Language Models

Abstract

The purpose of this paper is a philosophical exposition of the conceptual limitations and the corresponding epistemic value of applications of dialogic large language models (LLMs) based on the Transformer architecture as rational [co-]agents in the context of practical cognitively mediated human activity. To this end, the following research tasks are accomplished in the study: 1) a philosophical review of the fundamental intuitions, general nature and functional mechanics of advanced LLM implementations; 2) finding the limitations of the instrumental epistemic value of LLMs as products of a specific chain of machine learning procedures, which includes direct expert participation by humans. The practical significance of the study lies in the intuitive and compelling crystallization of the humanitarian understanding of LMMs through the theoretical and philosophical explication of their nature and design, which is grounded on the following new research results: LLMs are presented as a functionalist mechanistic project of statistical language and speech modeling as a model of knowledge as a semantic model of reality – 1) a model 2) of a model 3) of a model of reality; it is shown that due to the significant distance of reality mediation, intra-model connections lose in their factual capacity; іt is also demonstrated that an LLM is a product of machine learning of a certain linguistic behavior with a purpose and values that are radically different from those of human cognition, which makes the justifiability of their employment as components of any high-risk decision support systems to be fundamentally questionable.