The Propositional Turn in AI Knowledge Representation: Epistemic Accountability and the Limits of Sub-Propositional Symbols

This preprint presents a philosophical and architectural argument for what it calls the propositional turn in AI knowledge representation. The central claim is that knowledge-oriented AI systems should not treat tokens, entities, triples, embeddings, retrieved passages, or generated text as sufficient primary units of epistemic governance. Instead, when an AI system is expected to support verification, revision, provenance, explanation, and inferential accountability, these operations must be anchored to proposition-like commitments: bounded, truth-apt units that can be inspected, challenged, revised, and connected to other commitments. The paper first distinguishes the truth-aptness failure of words and entities from the boundary-indeterminacy problem of triples. It then argues that distributed representations and mechanistic explanations of model behavior do not by themselves provide an auditable layer of epistemic commitments. The paper develops a functional account of propositions as minimal adequacy units for epistemic governance, while avoiding a commitment to any specific metaphysics of propositions or to the Language of Thought hypothesis. It also situates the argument in relation to knowledge graphs, nanopublications, micropublications, semantic units, belief revision, mechanistic interpretability, and probabilist epistemology. The architectural consequence is sketched in terms of propositional semantic units: encapsulated proposition-like objects carrying normalized content, scope conditions, epistemic status, provenance, revision history, and typed inferential relations. The paper argues that such units provide a minimal governance layer for AI systems that must make claims inspectable, revisable, explainable, and accountable. This is a preprint version and has not yet undergone peer review.