The paper explains how workers' expectations of being discriminated against can be self-confirming, accounting for the persistence of unequal outcomes in the labour market even beyond the causes that originally generated them. The theoretical framework used is a two-stage game of incomplete information in which one employer promotes only one among two workers after having observed their productivity, which is used as a signal of their ability. Workers who expect to be discriminated against exert a lower effort on average, because of a lower expected return, thereby being promoted less frequently even by unbiased employers. This implies that achievements of minority groups may not improve when the fraction of discriminatory employers actually decreases, and such a mechanism is robust both to trial work periods and to affirmative actions like quotas.