We design a laboratory experiment to illuminate the channels through which relatively more attractive individuals receive higher wages. Specifically, we are able to distinguish taste-based discrimination from rational statistical discrimination and biased beliefs. Using three realistic worker tasks to increase the external validity of our results, we find that the “beauty premium” is highly task-specific: while relatively more attractive workers receive higher wage bids in a bargaining task, there is no such premium in either an analytical task or a data entry task. The premium in the bargaining task is driven by biased beliefs about worker performance. We find that there is substantial learning after worker- specific performance information is revealed, highlighting the importance of accounting for longer-run interactions in studies of discrimination.