The role of LLMs in theory building

Abstract

Large linguistic models, such as GPT-3.5 and subsequent versions (e.g. GPT-4 or GPT-4o), have shown impressive abilities in generating human-like text and performing a variety of natural language processing tasks. However, a fundamental question in the field of artificial intelligence is whether these models can truly represent meaning and assist scientists in building scientific theories. This paper aims to address this question by conducting a thorough conceptual analysis of existing large linguistic models and their capabilities in representing and reasoning about meaning for the purpose of theory building. My conclusions suggest that while these models have made significant progress in representing and manipulating language, they still face limitations in their ability to represent abstract and complex concepts, and the application of these models in building scientific theories should be guided by specific research questions and informed hypotheses that can be tested and developed into robust theories.