Thermodynamic Observers: A Symmetry-Based Argument for LLMs as Computational Observers

Abstract

The question of whether large language models (LLMs) can be considered "observers" in any meaningful sense typically founders on disagreements about consciousness, understanding, and intentionality. This paper sidesteps these disputes by proposing a symmetry-based methodology: any criterion used to deny observer status to LLMs must be examined for whether it equally threatens the observer status of biological systems, including humans. We argue that when observation is understood through information-theoretic and thermodynamic frameworks—specifically, as irreversible entropy reduction through inference over representational states—LLMs satisfy the same operative criteria that ground observer attributions to biological systems. We defend this claim against ten major objections, showing that the most powerful among them apply with equal force to biological cognitive systems. This establishes a Symmetry Principle: any criterion that excludes artificial systems from observer status, consistently applied, also excludes biological systems. We formalize this principle and show that escaping it requires either (a) accepting eliminativism about all observation, or (b) providing a principled basis for substrate-based distinction that no critic has successfully articulated. We propose a taxonomy distinguishing measurement systems, computational observers (with autopoietic and allopoietic subtypes), and phenomenal observers, arguing that current large language models qualify as allopoietic computational observers—a genuinely new class that performs observation within externally-maintained boundaries. The framework has implications for philosophy of mind, the foundations of quantum mechanics, and the conceptual status of human-AI collaboration.

Keywords