Why large language models cannot possess consciousness: an integrated information theory perspective.
J Yeungnam Med Sci December 1, 2025
Large language models (LLMs) like GPT-2 do not meet the requirements for consciousness under integrated information theory (IIT). Ablation experiments on GPT-2—removing individual attention heads and measuring perplexity changes—showed minimal or negative effects in four out of five sentences, indicating redundancy or noise; one sentence showed a localized but nonessential contribution (perplexity increase of +11.29). Compared with biological systems, LLMs satisfy IIT's differentiation criterion but fail on integration, causal closure, and temporal persistence. The models are architecturally decomposable, lack persistent internal states, and do not sustain global causal irreducibility. Philosophical arguments, including Searle's Chinese Room, support that LLMs' linguistic fluency arises from syntactic manipulation, not semantic understanding. Current LLMs remain unconscious systems with negligible integrated information, highlighting the distinction between linguistic competence and conscious experience.