This dissertation shows that the generalizations that speakers project from the lexical exceptions of their language are biased to be natural and output-oriented, and it offers a model of the grammar that derives these biases by encoding lexical exceptions in terms of lexically-specific rankings of universal constraints in Optimality Theory (Prince & Smolensky 1993/2004). In this model, lexical trends, i.e. the trends created by the phonological patterning of lexical exceptions, are incorporated into a grammar that applies deterministically to known items, and the same grammar applies stochastically to novel items. The model is based on the Recursive Constraint Demotion algorithm (Tesar & Smolensky 1998, 2000; Tesar 1998; Prince 2002), augmented with a mechanism of constraint cloning (Pater 2006, 2008b). Chapter 2 presents a study of Turkish voicing alternations, showing that speakers replicate the effects that place of articulation and phonological size have on the distribution of voicing alternations in the lexicon, yet speakers ignore the effects of vowel height and backness. This behavior is tied to the absence of regular effects of vowel quality on obstruent voicing cross-linguistically, arguing for a model that derives regular phonology and irregular phonology from the same universal set of OT constraints. Chapter 3 presents a study of Hebrew allomorph selection, where there is a trend for preferring the plural suffix [-ot] with stems that have [o] in them, which is analyzed as a markedness pressure. The analysis of the trend in terms of markedness, i.e. constraints on output forms, predicts that speakers look to the plural stem vowel in their choice of the plural suffix, and ignore the singular stem. Since real Hebrew stems that have [o] in the plural also have [o] in the singular, Hebrew speakers were taught artificial languages that paired the suffix [-ot] with stems that have [o] only in the singular or only in the plural. As predicted, speakers preferred the pairing of [-ot] with stems that have [o] in the plural, i.e. speakers prefer the surface-based, output-oriented generalization. Chapter 4 develops the formal theory of cloning and its general application to lexical trends, and explores its fit with the typologically available data. One necessary aspect of the theory is the "inside out" analysis of paradigms (Hayes 1999), where the underlying representations of roots are always taken to be identical to their surface base form, and abstract underlying representations are limited to affixes. An algorithm for learning the proposed underlying representations is presented in a general form and is applied to a range of test cases.