Usable Information Theory: A Framework for Meaning, Utility, and Intelligence in Bounded Agents

Abstract: Classical information theory measures syntactic uncertainty but deliberately excludes meaning, utility, and relevance. This paper proposes Usable Information Theory (UIT), a framework in which information becomes meaningful only through its capacity to produce predictive, compressive, or generative utility relative to a bounded agent and its world model. We define usable information UA(I, MA) = I − H(I | MA), semantic entropy ΣA(I, t) = H(I | MA(t)) as the residual information not yet integrated into the agent’s world-model, and informational free energy FA(I) = I − ΣA as extractable epistemic work. Intelligence is defined as the rate of semantic entropy reduction. We argue that apparent discontinuities in learning—semantic phase transitions—arise from continuous world-model updates cascading through reservoirs of latent high-ΣA information, reconciling UIT with the Free Energy Principle as nested levels of description. We position UIT against Shannon (1948), Carnap & Bar-Hillel (1953), Floridi (2011), Dretske (1981), Kolmogorov (1965), Tishby, Pereira & Bialek (1999), Schmidhuber (2010), and Kolchinsky & Wolpert (2018). UIT is presented as a falsifiable framework with explicit empirical predictions, including a proposed experimental test of the latent cascade mechanism in language models.