The Epistemic Anchor for LLMs Is Not the Dataset. It's the Governance Layer.

The AI governance and strategic communications communities have named data poisoning as a significant threat to LLM integrity. The response has largely focused on dataset hygiene: provenance tracking, source diversity, and adversarial content filtering. These controls are necessary but insufficient, because they misidentify the epistemic structure of the problem. LLMs have no internal epistemic anchor. They are probabilistic interpreters of text distributions, not truth-tracking systems. Datasets cannot serve as anchors because they are themselves contested, mutable, and attackable epistemic artifacts. This paper argues that the only viable epistemic anchor for LLMs is external to the model: a governance architecture that constrains how models learn, defines what outputs are admissible, and establishes institutional accountability for model behavior. The paper defines five components of this governance anchor—provenance governance, behavioral governance, interpretability governance, decision-boundary governance, and institutional accountability—and explains why accepting the externality of the anchor is the move that makes the data poisoning problem tractable. The epistemic anchor for LLMs is not their training data. It is the governance layer.