This paper analyzes statistical discrimination in a model with endogenous human capital formation and a frictionless labor market. It is shown that in the presence of two distinguishable but ex-ante identical groups of workers discrimination is sustainable as an equilibrium outcome. This is true irrespective of whether there are multiple equilibria when the groups have no distinguishable characteristics. When an affirmative action policy consisting of an employment quota is introduced in the model it is shown that affirmative action can "fail" in the sense that there may still be equilibria where the groups are treated differently. However, the incentives to invest for agents in the disadvantaged group are better in any equilibrium under affirmative action than in the most discriminatory equilibrium without the policy. Thus, the lower bound on the fraction of agents from the disadvantaged group who invest in their human capital is raised by the policy. The welfare effects are ambiguous. It is demonstrated that the policy may increase the incentives to invest and reduce the expected payoffs for all agents in the target group simultaneously. Indeed, the policy may hurt the intended beneficiaries even when the initial equilibrium is the worst equilibrium for the targeted group.