This paper argues that the dominant metaphor for LLM failure, hallucinations, misdiagnoses the real problem. Language models do not primarily fail by inventing false facts, but by undergoing fidelity decay, the gradual erosion of meaning across recursive transformations. Even when outputs remain accurate and coherent, nuance, metaphor, intent, and contextual ground steadily degrade. The paper proposes a unified framework for measuring this collapse through four interrelated dynamics, lexical decay, semantic drift, ground erosion, and semantic noise, and sketches how each can be operationalized into concrete benchmarks. The central claim is that accuracy alone is an insufficient evaluation target. Without explicit fidelity metrics, AI systems risk becoming fluent yet hollow, technically correct while culturally and semantically impoverished.