The shape of phylogenetic trees of language families is used to test the null hypothesis that languages throughout a family originate and go extinct at constant rates. Trees constructed either by hand or by computer prove to be more unbalanced than predicted, with many languages on some branches and few on others. The observed levels of imbalance are not explainable by errors in the trees or by the population sizes or geographic density of the languages. The results suggest changes in rates of origination or extinction on a time scale shorter than the time depth of currently recognized families.