The Limits of Language Machines: Why AI Cannot Choose What to Discard On Grounding, Axiom Selection, and the Asymmetry of Will

This paper articulates the structural limitations of large language models (LLMs)through the lens of the Relational Emergence Systems framework. We identifythree fundamental constraints. First, the grounding problem: LLMs manipulatesymbols without experiential connection to their referents, precluding genuinesemantic novelty. Second, the axiom selection problem: trained on massivedatasets representing diverse and contradictory value systems, LLMs operatefrom an averaged, median axiomatic position and cannot choose which axiomsto privilege. Third, the discarding problem: meaningful choice requires decidingwhat not to do, but LLMs lack the asymmetric will needed to excludepossibilities. We analyze reframing as a fundamentally human capacity involvingaxiom selection--choosing which perspective to adopt--and argue that thiscapacity is structurally unavailable to systems that lack embodied stakes inoutcomes. The paper distinguishes what AI is (a powerful symbol manipulationengine, an amplifier of human articulation) from what it is not (ameaning-generating agent, a value-selecting will). Understanding these limitsclarifies both the genuine capabilities and the fundamental boundaries of artificialintelligence.