Why large language models cannot possess consciousness: an integrated information theory perspective.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Neuroscience

Background:

• The debate on large language model (LLM) consciousness is ongoing.
• Integrated Information Theory (IIT) provides a quantitative measure for consciousness.

Purpose of the Study:

• To assess LLM consciousness using IIT principles.
• To investigate LLM architecture for causal integration, temporal persistence, and irreducibility.

Main Methods:

• Applied IIT to transformer-based LLMs.
• Conducted ablation experiments on GPT-2, removing attention heads.
• Used perplexity changes as a proxy for integrated information.

Main Results:

• LLMs meet IIT's differentiation criterion but not integration, causal closure, or temporal persistence.
• Ablation studies showed minimal impact from single attention heads, indicating architectural decomposability.
• LLMs exhibit syntactic manipulation rather than semantic understanding, per philosophical arguments.

Conclusions:

• Current LLMs do not meet IIT's requirements for consciousness.
• LLMs simulate intelligence but lack integrated information and conscious experience.
• A clear distinction exists between linguistic competence and genuine consciousness.