African Americans face shorter employment durations than apparently similar whites. We hypothesize that employers discriminate in either acquiring or acting on ability-relevant information. We construct a model in which firms may "monitor" workers. Monitoring black but not white workers is self-sustaining: new black hires are more likely to have been screened by a previous employer, causing firms to discriminate in monitoring. We confirm the model's prediction that the unemployment hazard is initially higher for blacks but converges to that for whites. Two additional predictions, lower lifetime incomes and longer unemployment durations for blacks, are known to be strongly empirically supported.