Apparent Algorithmic Bias and Algorithmic Learning

Paper by Anja Lambrecht and Catherine E. Tucker: “It is worrying to think that algorithms might discriminate against minority groups and reinforce existing inequality. Typically, such concerns have focused on the idea that the algorithm’s code could reflect bias, or the data that feeds the algorithm might lead the algorithm to produce uneven outcomes.

In this paper, we highlight another reason for why algorithms might appear biased against minority groups which is the length of time algorithms need to learn: if an algorithm has access to less data for particular groups, or accesses this data at differential speeds, it will produce differential outcomes, potentially disadvantaging minority groups.

Specifically, we revisit a classic study which documents that searches on Google for black names were more likely to return ads that highlighted the need for a criminal background check than searches for white names. We show that at least a partial explanation for this finding is that if consumer demand for a piece of information is low, an algorithm accumulates information at a lesser speed and thus takes longer to learn about consumer preferences. Since black names are less common, the algorithm learns about the quality of the underlying ad more slowly, and as a result an ad is more likely to persist for searches next to black names even if the algorithm judges the ad to be of low-quality. Therefore, the algorithm may be likely to show an ad — including an undesirable ad — in the context of searches for a disadvantaged group for a longer period of time.

We replicate this result using the context of religious affiliations and present evidence that ads targeted towards searches for religious groups persists for longer for groups that are less searched for. This suggests that the process of algorithmic learning can lead to differential outcomes across those whose characteristics are more common and those who are rarer in society….(More)”.