The neurocognitive mechanisms that support the generalization of semantic representations across different languages remain to be determined. Current psycholinguistic models propose that semantic representations are likely to overlap across languages, although there is evidence also to the contrary. Neuroimaging studies observed that brain activity patterns associated with the meaning of words may be similar across languages. However, the factors that mediate cross-language generalization of semantic representations are not known. We here identify a key factor: the depth of processing. Human participants were asked to process visual words as they underwent functional MRI. We found that, during shallow processing, multivariate pattern classifiers could decode the word semantic category within each language in putative substrates of the semantic network, but there was no evidence of cross-language generalization in the shallow processing context. By contrast, when the depth of processing was higher, significant cross-language generalization was observed in several regions, including inferior parietal, ventromedial, lateral temporal, and inferior frontal cortex. These results support the distributed-only view of semantic processing and favour models based on multiple semantic hubs. The results also have ramifications for psycholinguistic models of word processing such as the BIA+, which by default assumes non-selective access to both native and second languages.