Profits of Prejudiced Algorithms

Firms are starting to replace humans with algorithms in important screening decisions, but there are potential spillovers of human biases contained in datasets to subsequent algorithmic predictions. When these biases are motivated by human prejudices, there are risks of algorithms perpetuating discrimination. I prove that when datasets are generated by a sufficiently discriminatory human, firms are more profitable when training discriminatory algorithms. If instead enough affirmative action is instituted in favor of a disadvantaged group, firms are more profitable when training algorithms that inflate scores for this group, but this effect diminishes with excess affirmative action.