Tacit knowledge in large language models

Abstract

Recent debates about large language models (LLMs) question whether these systems merely reproduce training data or demonstrate genuine understanding. This paper challenges the “stochastic parrot” critique and argues that LLMs possess tacit knowledge, broadly defined as knowledge that is difficult or impossible to articulate explicitly. Tacit knowledge, however, has several features. LLMs exhibit two of the three forms of tacit knowledge: 1) knowledge that could theoretically be codified but is too costly to translate into explicit information, and 2) knowledge of nuance and subtext encoded in language, but not 3) embodied knowledge gained through sensory experience. These models learn latent information and unspoken rules from their training corpus, developing mental models that let them generalize to novel situations. However, LLMs still remain subject to Hayekian constraints of dispersed knowledge and bounded rationality.