In a controlled laboratory experiment we find that people are overconfident about the relative performance of their (randomly assigned) in-group on an intelligence test. Our design explicitly rules out any confounding effect of individual overconfidence in the self by excluding the self from the reference group. Although in-group overconfidence fades out once relevant information is provided, the speed of convergence is slow, because the belief-updating process is asymmetric (putting more weight on positive than on negative information about the in-group). Our results suggest that beliefs about group-level performance can be distorted in a similar ego-enhancing fashion as beliefs about individual performance. A bias in beliefs about group-level abilities can have important societal and economic consequences, e.g., for patterns of statistical discrimination in hiring contexts.