Two experiments test the naturalness hypothesis of velar palatalization. This hypothesis, based on surveys of various languages with velar palatalization, states that if a language has palatalization before [e], then it will have palatalization before [i], but not necessarily vice versa. Serbian is a prima facie counterexample to this generalization in certain morphosyntactic contexts, including the present-tense paradigm examined in this paper. In this context, Serbian palatalizes a velar stop [k] to a palatoalveolar affricate [ʧ] before [e] but not before [i]. Two experiments are conducted to test whether Serbian-speaking children and adults generalize from the existing pattern of palatalization before [e] to the natural pattern of palatalization before both mid and high vowels. The results from the first experiment show that children conform to the phonetically natural pattern but adults do not. These results suggest that speakers must be exposed to the pattern that "violates" the phonetically natural one for a substantial period of time before overwriting the phonetically natural pattern. The results from the second experiment, artificial pattern learning, show that the type of task and the type of palatalization (before [i] or [e]) play a crucial role, while age does not. These findings strengthen the hypothesis that subjects are more likely to choose a phonetically natural form presented to them than to volunteer it.