Models of bilingual speech production generally assume that translation equivalent lexical nodes share a common semantic representation. Though this type of architecture is highly desirable on both theoretical and empirical grounds, it could create difficulty at the point of lexical selection. If two translation equivalent lexical nodes are activated to roughly equal levels every time that their shared semantic representation becomes activated, the lexical selection mechanism should find it difficult to “decide” between the two (the “hard problem”) – yet in some cases bilinguals benefit from the presence of a translation equivalent “competitor”. In this article, we review three models that have been proposed as solutions to the hard problem. Each of these models has difficulty accounting for the full range of findings in the literature but we suggest that these shortcomings stem from their acceptance of the assumption that lexical selection is competitive. We argue that without this assumption each proposal is able to provide a full account of the empirical findings. We conclude by suggesting that the simplest of these proposals should be rejected before more complicated models are considered.