In an experimental study of language switching and selection, bilinguals named numerals in either their first or second language unpredictably. Response latencies (RTs) on switch trials (where the response language changed from the previous trial) were slower than on nonswitch trials. As predicted, the language-switching cost was consistently larger when switching to the dominant L 1 from the weaker L 2 than vice versa such that, on switch trials, L 1 responses were slower than in L 2 . This ''paradoxical'' asymmetry in the cost of switching languages is explained in terms of differences in relative strength of the bilingual's two languages and the involuntary persistence of the previous language set across an intended switch of language. Naming in the weaker language, L 2 , requires active inhibition or suppression of the stronger competitor language, L 1 ; the inhibition persists into the following (switch) trial in the form of ''negative priming'' of the L 1 lexicon as a whole.