This study compared the functioning of positively and negatively worded personality items using item response theory. In Study 1, word pairs from the Goldberg Adjective Checklist were analyzed using the Graded Response Model. Across subscales, negatively worded items produced comparatively higher difficulty and lower discrimination parameters than positively worded items and yielded almost no information. Model fit was examined for two forms of each scale: parameters freely estimated versus parameters estimated with item pairs constrained to be equal. Greater misfit was found in the latter. In Study 2, positively and negatively worded items from a more commonly formatted personality assessment were compared. Parameters again differed, albeit to a lesser extent, and model fit was improved in four out of five scales by the removal of negatively worded items. These results indicated that positively and negatively worded items were not psychometrically interchangeable and that negatively worded items have limited utility. Implications and future directions are discussed.