Abstract

Cognitive terms such as “understanding" or “reasoning" are increasingly applied to large language models, even by technically informed researchers. This paper argues that such applications are best understood neither as literal attributions nor as mere loose talk, but as instances of metaphor as conceptual adaptation: the creative extension of concepts beyond their original domain under pressure of what I call conceptual needs. Drawing on the Strawson–Kant tradition on imagination and concept-application, inferentialist semantics, Bermúdez's theory of rational framing, and Gentner's structure-mapping account of analogy, I develop a framework with four interlocking components. First, I argue that metaphorical extension selectively transfers inferential connections – not reference – from a source domain to a target domain. Second, I show that this transfer is guided by conceptual needs: mismatches between our conceptual repertoire and our circumstances that possess a distinctive aptic normativity (a normativity of fit). Third, I argue that whether the transferred connections hold in the target domain is empirically evaluable, and that mechanistic interpretability now furnishes the tools for this evaluation. Fourth, drawing on Bowdle and Gentner's "career of metaphor" model and Müller's sleeping/waking framework, I propose that cognitive terms applied to AI are undergoing a career – a dynamic trajectory from novel comparison through inferential negotiation towards potential conventionalisation. This framework makes headway beyond the debate over the literal applicability of cognitive terms to AI by replacing it with graduated, empirically tractable questions about the aptness of specific inferential transfers, and it reveals why the choice of cognitive vocabulary for AI is not merely terminological, but evaluatively and practically consequential.