Patients displaying mild symptoms of Alzheimer's disease sometimes have more difficulty naming items from an artifact than from a natural kind category; others displaying more severe symptoms almost always have more difficulty naming items from a natural kind than from an artifact category. This paper examined a computational model of this double dissociation (Devlin, Gonnerman, Andersen, & Seidenberg, 1998). Four basic tests of the model were proposed: The model should be able to generalize to new exemplars, the model should be expandable such that training sets of a realistic size can be used, the model's performance should not be unduly affected by small changes in architecture, and the learning algorithm should produce results that are not inconsistent with any major underlying factor of semantic organization. The model was found to be deficient in all four areas. Results reported from the model may therefore have been idiosyncratic to the model and not reflect general properties of a real semantic system.