The Pristine Fallacy: Why Chat Data Is Not a Clean Training Source

The Pristine Fallacy. Chat data is treated as human-written and therefore safe for model training. It is not. Three independent grounds: (1) AI-habituated writers produce text bearing mediation signatures even when writing unaided; (2) conversational feedback loops compress user inputs toward the model's distributional center; (3) single-model users develop model-specific accommodations that brand their inputs. The pristine fallacy is the assumption that data can be classified as clean or contaminated based on the identity of the writer rather than the conditions of production. Human-written describes the author, not the distribution. The contamination gradient (habituation depth × turn position × task type × model diversity) converts the thesis from polemic into audit framework. Four falsification conditions. Part of the Model Collapse Triptych with EA-MMRS-FIVESUB-01 and EA-LO-WOLFBOY-01. Series: Meaning Feudalism. SPXI-TLP v2.2 hardened. EA-SEM-PRISTINE-01 v1.0.