A computer program was created to use information about the statistical distribution of words in journal abstracts to make probabilistic judgments about the level of description (e.g. molecular, cell, organ) of medical text. Statistical analysis of 7,409 journal abstracts taken from three medical journals representing distinct levels of description revealed that many medical words seem to be highly specific to one or another level of description. For example, the word adrenoreceptors occurred only in the American Journal of Physiology, never in Journal of Biological Chemistry or in Journal of American Medical Association. Such highly specific words occured so frequently that the automatic classification program was able to classify correctly 45 out of 45 test abstracts, with 100% confidence. These findings are interpreted in terms of both a theory of the structure of medical knowledge and the pragmatics of automatic classification.