While the literature demonstrated that automation reduces employment in routine jobs (job polarization), its impact on wages is still unclear and the debate open. By applying Counterfactual Quantile Regressions to historical data, this paper analyzes the channels through which automation affected wage inequality in the U.S. labor market during the 1990s. Contrary to conventional wisdom, we find that the observed decline in wage inequality among low earners was not due to lower prices paid for technology-substitute occupational tasks, but instead due to more homogeneous wages of workers performing these tasks. This evidence is consistent with a model of directed (routine-biased) technical change in which skill-heterogeneous workers face endogenous occupational choices and learning costs in connection with operating new technology. In this model, directed technical change reduces wage inequality among low earners by shrinking the skill distribution of routine workers, thus making their wages more homogenous as observed in data.