The present study explores learning vowel harmony with exceptions using the artificial language learning paradigm. Participants were exposed to a back/round vowel harmony pattern in which one affix (either prefix or suffix, depending on the condition) alternated between /me/ and /mo/ depending on the phonetic feature of the stem vowels. In Experiment 1, participants were able to learn the behaviors of both alternating and non-alternating affixes, but were more likely to generalize to novel affixes for non-alternating items than alternating items. In Experiment 2, participants were exposed to training data that contained non-alternating affixes in prefix position while alternating affixes were all suffixes, or vice versa. Participants were able to extend the non-alternating affixes to the novel direction, suggesting that participants inferred a non-directional harmony pattern. Overall, the patterns of alternating affixes are harder to learn than patterns of exceptions that do not alternate, which aligns with previous findings supporting a non-alternation bias. Our study raises the question of how biases towards exceptionality and directionality interact in phonological learning.