The Swim Instructor Who Never Swam: _Two Kinds of Competence in Artificial Intelligence_

The debate over whether artificial intelligence systems "understand" language has reached an unproductive impasse, with skeptics denying any philosophically significant competence and enthusiasts implying capacities far beyond what current systems possess. This paper argues that the impasse results in part from a persistent conflation of two different competence profiles. We define Textbook Competence as the capacity to perform context-sensitive, inferentially disciplined transformations on structured symbolic material across registers and tasks without requiring direct immersion in the domain described. Steeped Competence is then defined as the capacity acquired through sustained participation in a domain, shaped by perceptual, motor, social, and cultural attunement that is often difficult to verbalize and resistant to transmission through language alone. Current large language models plausibly exhibit the former on an operational reading and lack the latter in the ordinary human sense. We position this binary relative to more refined cuts in the recent literature and argue that it tracks a real categorical break in acquisition mode whose operational consequences the more refined cuts are not designed to make perspicuous. The brittleness profile documented in recent empirical work — strong performance on tasks statistically near the training distribution, sharp degradation under distribution shift — is precisely what the framework predicts and is therefore positive evidence for it, not merely a caveat on the empirical claim. By distinguishing these two competence profiles and mapping the evidence for each, we clarify the apparent paradox of systems that seem to "understand everything and nothing." The distinction is offered as an analytic tool with substantive philosophical and empirical commitments, not as a complete taxonomy of cognition.