Imagine walking through the forest when a small, brown creature darts across your path. Do you keep walking or back away? Your reaction depends on whether you think an animal with the glimpsed features

Imagine walking through the forest when a small, brown creature darts across your path. Do you keep walking or back away? Your reaction depends on whether you think an animal with the glimpsed features is likely to be dangerous. The need to predict the likelihood of an unknown property (dangerous) given some known properties (small and brown) is something that confronts us on a daily basis. Past experience, in the form of knowledge about categories, can be used to guide such predictions. For example, if the small, brown creature is categorized as a rodent, knowing the typical properties of rodents helps to assess the danger posed by the unknown animal. Support for inductive prediction is thought to be one of the central functions of categories (Anderson, 1991). The use of categories in making predictions is complicated by the fact that category membership is often uncertain. The creature might be a rodent, but given its known properties it might instead be a snake or a bird. The alternative categories differ considerably in their association with the property dangerous, so that the decision to exercise caution depends on the category. There are two approaches people can take when categorization is uncertain. Suppose you were to meet a person at a party who asks if you want to have coffee during the week. This person is talkative and outgoing . . . perhaps a little too outgoing. In all likelihood, these traits indicate that the person is friendly, in which case you would be interested in a coffee date. If, on the other hand, these traits are signs that the person is overbearing, you would be less interested. A prediction must be made (Will the person be an enjoyable coffee companion?) before categorization is certain. With a single-category approach, you choose the most likely category and use that to make your prediction. That is, if you decide the person is probably just friendly, then you act toward him or her as you would toward any friendly person, and the possibility that the person might be overbearing does not influence your prediction. With a multiple-category approach, on the other hand, you consider all possible categorizations when making your prediction. If there is some chance that the person is overbearing, you will be less likely to agree to have coffee than you would be if you were absolutely sure that the person is only friendly. That is, your prediction would reflect both possible categories. The multiple-category approach conforms to a classically rational view of decision making in which people attempt to maximize accuracy by using what amounts to a naive form of statistics. Many models of inductive reasoning are based on the idea that people consider multiple categories during the decision process. However, a body of empirical evidence suggests that when predicting unknown properties as described above, people generally adopt a single-category approach. In the real world, property prediction takes many different forms, and it may be that different prediction tasks lead people to take different approaches to prediction. In the present study, we were interested in whether people use cate-