Prompt ̸ = Knowledge: On the Epistemic Limits of In-Context Control in Large Language Models

Prompting has become the primary mechanism for interacting with large language models (LLMs), often accompanied by the implicit assumption that prompts can supply, inject, or constitute knowledge. We challenge this assumption by arguing that prompts do not generate, update, or instantiate knowledge within LLMs; rather, they function as transient control signals that shape inference trajectories grounded in pre-existing parameterized structures [2, 3]. Drawing on epistemology, control theory, and information theory [8, 9], we show that prompt-induced competence systematically fails to meet the criteria for knowledge attribution: persistence across contexts, generalization to novel situations, and counterfactual stability. We further explain why prompting appears to confer knowledge, tracing this to in-context learning effects, linguistic coherence, and human interpretive biases. The paper introduces the PAK(Prompt ̸= Knowledge) thesis, situates it within the DIKCA(DataInformation-Knowledge-Control-Architecture) framework [1, 6], and discusses implications for evaluation, governance, and AI system design.