What the Machine Already Knows. Parametric Knowledge, the Limits of Citation Optimization, and the Question of Who Writes the Canon of Language Models

Abstract Large language models answer from two sources: information retrieved at answer time and placed in the context window, and parametric knowledge, the information encoded in the model’s weights during training. The industry that optimizes visibility in AI answers has concentrated almost entirely on the first source. A study published by the agency Graphite in October 2026 argues that this focus misses the decisive instance: parametric knowledge shapes the search queries a model generates, the way it uses retrieved pages, and every answer produced without retrieval. This paper examines that argument, separates what the study’s data establish from what they suggest, and relates the findings to research on context-memory conflict and on data poisoning. It then asks a question the study does not pose: what follows for a public whose access to information is increasingly mediated by systems whose default assumptions are formed in corpora nobody outside the laboratories can inspect. Three structural problems are identified: the non-auditability of parametric knowledge, the absence of a workable correction procedure, and the institutional asymmetry created by training-data licensing. The paper proposes a verifiable architecture of source registers, correction procedures, and measurement rights, states its central claims as hypotheses with explicit refutation conditions, and gives a measurement note for practitioners and researchers. Author: Dirk de Pol, Omegacene.com, Berlin