Simulation Without Understanding: A Cognitive Risk Analysis of LLMs in Sensitive Human Environments

This paper critiques the illusion of understanding in large language models (LLMs), particularly those deployed as agentic AI systems. While such models appear fluent and responsive, they often simulate coherence without internal cognitive anchors, introspective audit layers, or epistemic memory. The result is an artificial fluency that conceals semantic drift, output inconsistency, and shallow alignment. Drawing from architectural evaluation, cognitive science, and system dynamics, the paper introduces a multi-level audit approach that exposes these failures. It proposes a redirection from output performance metrics toward meaning fidelity, memory transparency, and belief-continuity frameworks. The argument is not against simulation, but against its use without cognitive accountability. This research challenges prevailing assumptions in AI deployment and suggests an urgent reevaluation of alignment, trust, and the very definition of “understanding” in machine system