The empirical employer learning literature \u85nds support for statistical discrimination using schooling. How economically relevant statistical discrimination (and consequently job market signaling) is depends on how fast \u85rms learn about workersproductive types. I show that \u85rms learn quickly: on average expectations errors about productivity decline by 50 percent within three years. Based on this I \u85nd that signaling contributes at most 25 percent to the gains from schooling. Moreover, the standard human capital model has similar implications for earnings as the model of employer learning and is consistent with variation in on-the-job training by schooling and ability. *) I would like to thank my committee chairman, Robert Topel...