Manuel Boissenin, France

With Large Language Models (LLMs) exhibiting astounding abilities in human language processing and generation, a crucial debate has emerged: do they truly understand what they process and can they be conscious? While the nature of consciousness remains elusive, this synthetic article sheds light on its subjective aspect as well as some aspects of their understanding. Indeed, it can be shown, under specific conditions, that a cognitive system does not have any subjective consciousness. To this purpose the principle of a proof, based on a variation of the thought experiment of the Chinese Room from John Searl, will be developed. The demonstration will be made on a transformer architecture-based language model, however, it could be carried out and extended to many kind of cognitive systems with known architecture and functioning. The main conclusions are that while transformers architecture-based LLMs lack subjective consciousness based, in a nutshell, on the absence of a central subject, they exhibit a form of “asubjective phenomenal understanding” demonstrably through various tasks and tests. This opens a new perspective on the nature of understanding itself that can be uncoupled with any subjective experience.

Language models, transformers, subjective consciousness, understanding, asubjectivity