A current area of controversy within the literature on visual word recognition concerns the extent to which semantic information influences the computation of phonology. Experiment 1 revealed that both the imageability effect and the ambiguity advantage seen in the speeded naming task are confined to words with atypical mappings between spelling and sound. Nonetheless, it is possible that either of these effects may arise from the operation of the direct rather than the semantic pathway. Experiment 2 therefore included nonword fillers in order to minimize semantic reliance during speeded naming. This manipulation removed the imageability effect, indicating a semantic locus, but the ambiguity advantage remained, suggesting a nonsemantic locus. These results are considered in the context of computational models that incorporate a semantic level of representation.