Algorithmic Bias?


Speaker


Abstract

The popular debate on “algorithmic bias” has largely focused on settings and mechanisms where an algorithm's decision making is for the most part isolated from direct interactions with other economic actors. An algorithmic outcome that appears biased has then often been attributed to a bias being hard-coded or the idea that algorithms can learn bias from the data that itself reflects bias. However, many algorithms operate in consumer markets where their outcomes are influenced by market forces such as demand and competition. We present data from two field tests that demonstrate cases of apparent algorithmic bias as an outcome of such market forces. First, we explore data from a field test of how an algorithm delivered ads promoting job opportunities in the Science, Technology, Engineering and Math (STEM) fields. This ad was explicitly intended to be gender-neutral in its delivery. Empirically, however, fewer women saw the ad than men. This happened because younger women are a prized demographic and are more expensive to show ads to. An algorithm which simply optimizes cost-effectiveness in ad delivery will deliver ads that were intended to be gender-neutral in an apparently discriminatory way, due to crowding out. This empirical regularity extends to other major digital platforms. Second, we demonstrate in the context of search advertising that if consumer demand for a piece of information is low, an algorithm may take longer to learn about consumer preferences, leading to differential outcomes across those whose characteristics are more common and those who are rarer in society. As a result, the algorithm may be likely to show an undesirable ad to a disadvantaged group for a longer time period as it takes longer for the algorithm to learn about the lack of efficacy of the ad. Together, both studies demonstrate instances where uneven outcomes are linked to algorithmic decision making. However, rather than being a direct result of the algorithm, they are an outcome of, in the first case, spillover effects between different economic actors and, in the second, of different rates of algorithmic learning.