From Tabula Rasa to Inductive Bias: Reframing Locke's Problem in the Age of Generative AI

Abstract

Large language models (LLMs) often appear to vindicate a radical empiricist picture: train on vast corpora of experience-like text, and capacities emerge without explicit symbolic rules. Yet contemporary machine learning research repeatedly emphasizes that what is learned, how quickly it is learned, and how well it generalizes depend crucially on prior constraints: architectural structure, training objectives, optimization dynamics, and representational bottlenecks. These constraints constitute inductive biases in a precise, technical sense. This paper develops a philosophical argument that uses LLMs as a case study to reassess the classical tabula rasa thesis associated with Locke and its descendants. I defend two claims. First, even the most “data-driven” generative models are saturated with structural priors that make learning possible at scale; thus, their success cannot be straightforwardly read as a triumph of unstructured empiricism. Second, once this is appreciated, the rhetorical appeal of a ``blank slate'' conception of the infant mind weakens further: if artificial systems trained on orders of magnitude more linguistic input than any child still require rich inductive biases, it is implausible that human cognition begins wholly unstructured. I then show how contemporary debates about whether LLMs “understand” language recapitulate the older dispute about whether experience alone can generate semantics and conceptual structure. The upshot is a more balanced, non-caricatured empiricism: experience matters, but explanation must explicitly account for the learning system's prior structure.