Decision tree learning algorithms produce accurate models that can be interpreted by domain experts. However, these algorithms are known to be unstable --- they can produce drastically different hypotheses from training sets that differ just slightly. This instability undermines the objective of extracting knowledge from the trees. In this paper, we study the instability of the C4.5 decision tree learner in the context of active learning. We introduce a new measure of decision tree stability, and define three aspects of active learning stability. Several existing active learning methods that use C4.5 as a component are compared empirically; it is determined that query-by-bagging yields trees that are more stable and accurate than those produced by competing methods. Also, an alternative splitting criterion, DKM, is found to improve the stability and accuracy of C4.5 in the active learning setting.