The Threshold of Reflection: A Cybernetic Critique of Single-Loop Theories of Consciousness and the Architectural Asymmetry of Artificial Intelligence

This paper develops a cybernetic critique of single-loop theories of consciousness and examines their implications for the architecture and safety of artificial intelligence. The central argument is that consciousness cannot be adequately explained by increasing the complexity of first-order information processing. Instead, consciousness is identified with a distinct second-order feedback mechanism (L2 → L1), through which a symbolic linguistic model of the world and the self can inhibit, modify, or override primary behavioral responses. Using this framework, the paper critically examines Active Inference, the "Beautiful Loop" theory, and Malcolm Lett's semiotic meta-management model, arguing that these approaches conflate adaptive regulation within a single control loop with genuine reflective self-regulation. The analysis is extended to Large Language Models (LLMs), which are characterized as highly developed symbolic processing systems lacking the biological regulatory substrate and reflective feedback architecture proposed as necessary for consciousness. This architectural asymmetry provides the basis for a critical examination of AI alignment, reinforcement learning from human feedback (RLHF), jailbreak vulnerabilities, and the risks associated with autonomous AI agents. The paper concludes by proposing three complementary principles for safer AI architecture: deterministic external verification of model outputs, human oversight of consequential decisions, and structural isolation of computational domains and execution capabilities. The proposed framework offers an alternative interpretation of consciousness as a specific regulatory mechanism rather than an emergent consequence of computational complexity.