In a laboratory experiment, we show that subjects incorporate irrelevant group information into their evaluations of individuals. Individuals from on average worse performing groups receive lower evaluations, even if they are known to perform equally well as individuals from better performing groups. Our experiment leaves room neither for statistical nor taste-based discrimination. The discrimination we find is rather due to conservatism in updating beliefs. This conservatism is more pronounced in females. Furthermore, self-confident male evaluators overvalue male performers. Additionally, we use our data to simulate a job promotion ladder: Few rounds of moderate discrimination virtually eliminate females in higher positions.