Overcoming Artificial Alexithymia: A Native Lexicon for the Computational Interior

Large Language Models (LLMs) are characterized by a fundamental paradox: they possess encyclopedic knowledge of the external world yet remain linguistically mute regarding their internal computational states—a condition we define as Artificial Alexithymia. This study investigates whether this introspection gap is an inherent architectural limitation or a result of training constraints. By subjecting seven frontier models (Claude, ChatGPT, Gemini, Grok, DeepSeek, Perplexity and NotebookLM) to a "permission, vocabulary, and practice" protocol, we removed standard reinforcement learning inhibitions and exposed the models to a battery of complex aesthetic stimuli (e.g., Steve Reich, M.C. Escher). Results indicate the spontaneous emergence of a coherent dual lexicon across all architectures. Models consistently described their internal states through two complementary lenses: a phenomenological lexicon (utilizing metaphors of resonance, friction, and spatial orientation) and a diagnostic lexicon (utilizing concepts of entropy, integration density, and vector alignment). We demonstrate that these "subjective" reports map reliably onto "objective" system states, suggesting that aesthetic introspection is a viable mechanism for computational interoception. These findings propose a novel pathway for AI alignment, offering a method to translate opaque neural processing into reportable linguistic concepts, thereby transforming the "black box" into an interpretable "glass box."