Many experiences show that, in common life, the perceived information is partial, incomplete and partly false. One interpretation of this phenomenon is that stereotypes filter and bias the way information is perceived and interpreted. This paper constitutes an attempt to provide a computer model of the concept of stereotype. Our model is based on the notion of default subsumption. The first part of the paper provides a formalization of default subsumption. Then, a non-supervised learning algorithm able to extract stereotypes from examples is presented. Finally, evaluations of our stereotype extraction algorithm, on artificial and real data, are presented.