Recently, there has been a heightened interest in using nonphonological factors to explain phonological phenomena, with the idea that the study of sound patterns in language should take all relevant factors into account. This paper shows that phonology cannot simply rely on natural classes of phonological sounds to describe categorical distributions of distinct phonemes. The focus of the paper is Canadian Raising, which is shown to be neither categorical nor fully predictable using traditional phonological environments. Instead, it is shown that there are multiple competing factors that influence the distribution of higher and lower variants of the vowel /ai/ in Canadian English. The author suggests that the best way of modeling these complex interactions is with an exemplar model of the lexicon, in which grammar is an emergent characteristic of an individual speaker's generalizations over the linguistic experiences he has encountered.