Abstract. Categorization is a fundamental human skill, and an important topic in AI and machine learning. Similarity-based categorization algorithms are among the most successful in human and machine classification. This work investigates principles underlying human categorization, focussing on people’s sensitivity to within-category feature correlations in perceptual categorization. To this end, we test two mathematical similarity measures against human categorization performance: the Mahalanobis similarity metric compared to the Euclidean similarity metric. The results from computational predictive model testing with data from two experiments show convincingly that the Euclidean similarity represents the psychology of categorization best.