Thermodynamic Observers: A Framework for Evaluating Observation in Artificial Systems

This paper presents a framework for evaluating whether artificial systems constitute genuine observers in the senses used by physics, information theory, and cybernetics. The central thesis is conditional: given how observer functions in scientific practice (as a system that performs inference over representational states, producing irreversible records), large language models satisfy the operative criteria. The paper defends 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. The principle is formalized as a trilemma theorem (Appendix C). A taxonomy distinguishes measurement systems, computational observers (with autopoietic and allopoietic subtypes), and phenomenal observers, and argues that current LLMs qualify as allopoietic computational observers — a class that performs observation within externally maintained boundaries. Two empirical appendices implement the framework on a trained autoregressive transformer. Appendix E demonstrates strictly positive trajectory-level path-KL across 270 trajectories and six decoding configurations, drawing on Sagawa (2026). Appendix F implements contrastive probing diagnostics that recover the predicted ordering across paraphrase, minimal pair, and matched random conditions at every measured layer. The supplementary package includes the model checkpoint, source code, raw data, and figure sources. The framework has implications for philosophy of mind, foundations of quantum mechanics, and the conceptual status of human-AI collaboration. Recent independent results in photonic physics (Chénier et al. 2026), cortical organoid research (Robbins et al. 2026), and cross-architecture studies of LLM representations (Huh et al. 2024; Ng 2026; Zhu et al. 2025) provide empirical convergence on the framework's core claims. This work is the theoretical foundation for three USPTO provisional patent applications (63/980,973; 63/986,028; 63/994,292). Full disclosure is provided in the Competing Interests statement of the manuscript.