In industrialised countries, men and women tend to work in different occupations. Particularly striking is the degree to which gender segregation has persisted despite changes in legislation, employers' attitudes and agents preferences. This paper presents a possible explanation of such persistence when agents have imperfect information about their probability of success in different occupations. Agents have prior beliefs about these probabilities, on which they base their choice of occupation and effort. They then update their belief according to Bayes' rule and transmit this posterior belief to their offspring. A feature of Bayesian learning models is that the distribution of long-rung beliefs depends on the initial distribution. Hence, if men and women of previous generations had different preferences over types of jobs, this would have affected their learning and hence the beliefs of the current generation. As a result, even if men and women become identical in their preferences and abilities, they hold different beliefs and hence choose different career paths, thus maintaining past segregation.