The relationships of adult butterflies were evaluated using the techniques of numerical taxonomy and various sets of characters. Using a total of 196 characters of the external and internal anatomy produced phenograms quite similar to the one arrived at by classical taxonomic techniques. There were, however, significant differences (e.g., relatively low similarity of the Papilionidae and Pieridae). Relationships based on most subsets of characters showed a “family resemblance” to one another and to the classical relationships. Relationships based on 96 internal and 100 external characters were compared by a randomization test. When evaluated by correlation coefficients the internal and external relationships were significantly different. When evaluated by distances they were not. Our results in general seem to support a “weak” version of the nonspecificity hypothesis—analysis of different large samples of characters ordinarily will yield positively correlated but non-identical patterns of relationship. This study has led to an appraisal of ideas of general classification and the utility of numerical taxonomy. Theoretical considerations make it seem unlikely that the idea of “overall similarity” has any validity. In any case it is clear that overall similarity cannot be dealt with operationally. All classifications are inherently special, and no one special classification is any more or less “correct” than any other. It is suggested that numerical taxonomy is a more sophisticated tool than is required for doing most routine taxonomic jobs. Numerical taxonomy, however, can be used in those situations where refined answers seem to be desirable, either where economic problems are involved or where light can be shed on general problems of population biology.