The traditional laboratory task of supervised classification learning tends to produce sensitivity only to information that discriminates between competing categories. Recent research broadening the study of category learning (e.g., inference learning) suggests that learners can be sensitive to structure beyond what is diagnostic. In previous work we interpreted these findings to reflect a generative (as opposed to discriminative) mode of learning and extended the phenomena to supervised observational learning of categories varying along two continuous dimensions. We now aim to demonstrate generative observational learning using binary dimensions. Categories were based on a uni-dimensional rule with within-category regularities in the form of a family resemblance structure (Experiment 1) or a perfect correlation between features (Experiment 2). Compared to classification learners, observation learners showed greater knowledge of both types of non-diagnostic structure. These results hold implications for how categories are learned and present a challenge to models grounded in discriminative category learning.