The predominant explanation for difficulty naming objects in Alzheimer's disease (AD) is impaired semantic memory. Two classes of findings challenge this. The lower the visual quality of the stimulus the less likely AD patients are to name it, suggesting a deficit of visual perception. The lower the name frequency, the less likely AD patients are to name an object, suggesting a deficit of lexical access. A mechanistic explanation is given for why a semantic memory deficit is sufficient to account for this range of data, provided components underlying task performance are interactive. Interactive parallel distributed processing networks were trained to associate visual patterns with semantic and lexical patterns. When semantic units were lesioned, networks were more sensitive to impoverished visual inputs. Networks also made a disproportionate number of errors to items trained with lower frequency and benefited from phonemic cues.