Three studies examined how people make feature inferences about exemplars whose category membership is uncertain. Participants studied categorized exemplars, were given a feature of a novel item and asked to make predictions about other features. Stimuli were constructed so that different inference strategies led to divergent feature predictions. Experiments 1 and 3 found that most participants used a feature association strategy where predictions were based on comparisons with exemplars similar to the test item. Experiment 2 showed that the dominance of feature association over categorical approaches to reasoning was not an artifact of stimulus complexity.