This paper is concerned with two issues: the nature of the methodology employed in the construction of language understanding systems within Artificial Intelligence (AI); and, secondly, the status of the “semantic primitives” employed within certain of those systems. On the first, it is argued that attempts to justify the methodology of AI research on natural language by appeal to the methods of the sciences are in general misguided. On the second issue, the paper argues, more specifically, that semantic primitives cannot be justified in the way that theoretical objects in the sciences (such as neutrinos) are. It suggests that such primitives are not essentially different from the surface words whose meanings they are used to express, and that recognition of this fact in no way limits their usefulness in linguistic research.