Large Language Models and the Patterns of Human Language Use

Abstract

Text-generating large language models (LLMs) such as ChatGPT are deep learning architectures that have been trained on immense amounts of text. Their ability to produce human-like text has led to claims that LLMs either possess or simulate some form of conscious experience and understanding, whereas others contend that LLMs just parrot human language. Both kinds of claims, however, render enigmatic why LLMs frequently produce text that is not only meaningful to humans but also a good response to a prompt. Against both claims, this chapter argues that the impressive degree to which stochastic patterns can be computationally reassembled into text that makes sense to humans can reveal the surprising extent to which human language use gives rise to and is guided by patterns. LLMs can generate language that appears similar to that of humans by means of stochastic patterns that model aspects of human language use. LLMs are trained not simply on language, but language use, and language use is intertwined with experiential patterns. By predicting patterns of language use, LLMs can also evoke phenomenological patterns, including clichés and biases. Our approach enables not only a new way of looking at the capabilities of LLMs, but also a reconsideration of the question of language use, sense-making, and their relation to the phenomenology of experience.

Topics & Concepts