How Do Disparities Reproduce Themselves? “Ground Truth” Inference From Utility-Maximizing Agent’s Sampling Behavior

• In Event: Learning About Others From Others: Implications for Stereotype Formation and Intervention

Abstract

Why do people often reproduce disparities they see rather than rectifying them? For instance, while mounting evidence suggests that the police stop and search African Americans more often than other races (Hetey et al., 2016; Pierson et al., 2020), such disparities are still prevalent and sometimes justified ("Maybe they are more prone to crime?"). Social psychologists (Hetey & Eberhardt, 2014, 2018) hypothesized that people with stereotypes linking Blacks to crime become more fearful of crime when seeing the overrepresentation of Blacks in the criminal justice system and are therefore more likely to uphold current practices. Will doing away with stereotypes end disparities? We argue not. If a learner believes an officer knows each race's "true crime rate" and maximizes expected utilities (expected rewards of catching criminals minus checking costs), they should infer that groups checked more often have higher crime rates and will follow the officer’s footsteps as a result of this Naive Utility Calculus ("NUC", Jara-Ettinger et al., 2016) rather than pre-existing stereotypes.

To test our hypothesis, we designed a “Golden Ticket" game, where a robot chicken lays eggs that may or may not have golden tickets but each egg always costs a token to buy. Player Alex knows each chicken's hit rate and does not waste tokens. Participants watched Alex play 12 chickens, each laying 6 eggs, which Alex could either buy or forgo. Afterwards, participants were asked to estimate each chicken’s hit rate and decide on whether to buy an egg.

In Experiment 1, Alex never opened the eggs so participants (62 U.S. adults recruited on Amazon Mechanical Turk) only knew the "check rates" (the probability of Alex buying an egg from a chicken). According to the NUC, the higher a chicken's hit rate, the higher the expected utility of buying from it, and thus the higher the check rate. We found that participants were indeed more likely to buy from chickens favored by Alex and this NUC model closely captured their hit rate inferences, suggesting that they were not just imitating Alex but also reasoning about the underlying motivation.

In Experiment 2, Alex opened the eggs they bought. Participants (67 U.S. adults recruited on MTurk) may solely rely on this new hit rate information ("Hit Rate Only"), disregard it ("Check Rate Only"), or combine it with check rates ("Check + Hit Rates") to infer population hit rates. The "Check Rate Only" model did not resemble human inferences whereas the other two models together captured most patterns. When combining information, participants weighted sample hit rates more heavily than check rates.

In summary, we found in Experiment 1 that people reproduced observed disparities without existing stereotypes: The more a knowledgeable, utility-maximizing agent sampled from a population, the higher people inferred the population hit rate to be. In Experiment 2, we changed people’s reliance on the agent’s sampling behavior by showing them hit rates in the samples. Our work provided a novel explanation for why disparities are persistent despite people's awareness and a potential intervention to combat this tendency.

Authors

• Yuan Meng, University of California - Berkeley
  Presenting Author

• Fei Xu, University of California - Berkeley
  Non-Presenting Author