Are Large Language Models Conscious?

This work evaluates the possibility of a conscious large language model (LLM) under the parameters of Integrated Information Theory (IIT). First, we examine the methodology behind IIT as a scientific theory of consciousness. Second, we detail the system requirements for consciousness under IIT. Third, we examine the basic architecture of modern LLMs. Finally, we subject the network architecture of current LLMs to the system requirements of IIT, evaluating each “physical postulate” IIT espouses individually. We conclude that LLMs meet none of IIT’s requirements for a physical system to be conscious, and briefly discuss the implications for future research.