Algorithmic Bias in Service

Abstract

Research shows that algorithms using sociodemographic data (e.g., race, gender, education, etc.) can produce biased outcomes that cause many consumers to be excluded from or endure lower levels of service. Though research suggests that these algorithms are more profitable than unbiased algorithms that do not use sociodemographic data, prior findings do not consider potential social effects of these algorithms on consumer demand. This research investigates the dynamic outcomes of competition between biased and unbiased algorithms in a market where word-of-mouth influences consumer choice behavior. Relative to unbiased algorithms, this research demonstrates that biased algorithms can be more profitable in the short run but less profitable in the long run, due to consumer word-of-mouth. Models and simulations show that word-of-mouth leads marginalized consumers to gravitate towards easier-to-access unbiased algorithmic services. Non-marginalized consumers, on the other hand, learn they have a relatively easier time accessing services anywhere. When sufficient numbers of marginalized and non-marginalized consumers learn from each other via word-of-mouth, long run demand is greater for unbiased algorithmic services. This research demonstrates that firms that use unbiased algorithms and account for social effects (e.g., word-of-mouth) in the algorithm’s design can reduce algorithmic bias while improving both long-term profits and societal well-being.

Keywords: algorithms, algorithmic bias, algorithmic fairness, discrimination, word of mouth, agent-based modeling