The Classification Error of Modern AI

This working paper proposes a system-theoretical classification of modern artificial intelligence systems. It distinguishes between cybernetically validatable tools, probabilistic transformation systems, and autonomous experience-capable intelligent systems. The central argument is that Large Language Models should not be classified as autonomous intelligent systems solely on the basis of linguistic performance. Instead, they are examined as probabilistic transformation systems whose outputs require external validation when truth, responsibility, or operational reliability are relevant. The paper introduces formal criteria for forward testability, backward testability, validation, error integration, and responsibility localization. It further discusses falsification conditions under which the proposed classification would require revision. The work is intended as a conceptual, falsifiable, and revisable framework for future empirical, technical, and governance-oriented research.