Home – Home – Jornals – Philosophy of Sciences 2025 number S5

This paper examines the problem of semantics of intensional contexts from a special perspective related to the principles of operation of generative artificial intelligence (AI). Intensional contexts such as beliefs, desires, knowledge and convictions pose a particular challenge for modern language models, since they require taking into account the meanings of expressions in logical possible worlds or cognitive states of subjects. This paper is a kind of sketch for the formulation of a problem that has an engineering and philosophical dimension. Here, we analyze the mechanisms of operation of transformers that use contextual embeddings to model the meanings of words through self-attention. It was revealed that modern language models are able to effectively process anaphoric dependencies and contextual connections, but they face limitations when interpreting intensional constructions. Particular attention is paid to experiments with vector representations of objects in multidimensional space, during which difficulties arise when distinguishing between subjective beliefs and objective reality. The nature of the difficulties indicates that working with intensional contexts requires not only a simple analysis of probabilistic connections between words, but also a deeper understanding of the semantics of linguistic expressions.