Working Paper: NBER ID: w30981
Authors: Amanda Y. Agan; Diag Davenport; Jens Ludwig; Sendhil Mullainathan
Abstract: Consumer choices are increasingly mediated by algorithms, which use data on those past choices to infer consumer preferences and then curate future choice sets. Behavioral economics suggests one reason these algorithms so often fail: choices can systematically deviate from preferences. For example, research shows that prejudice can arise not just from preferences and beliefs, but also from the context in which people choose. When people behave automatically, biases creep in; snap decisions are typically more prejudiced than slow, deliberate ones, and can lead to behaviors that users themselves do not consciously want or intend. As a result, algorithms trained on automatic behaviors can misunderstand the prejudice of users: the more automatic the behavior, the greater the error. We empirically test these ideas in a lab experiment, and find that more automatic behavior does indeed seem to lead to more biased algorithms. We then explore the large-scale consequences of this idea by carrying out algorithmic audits of Facebook in its two biggest markets, the US and India, focusing on two algorithms that differ in how users engage with them: News Feed (people interact with friends' posts fairly automatically) and People You May Know (people choose friends fairly deliberately). We find significant out-group bias in the News Feed algorithm (e.g., whites are less likely to be shown Black friends' posts, and Muslims less likely to be shown Hindu friends' posts), but no detectable bias in the PYMK algorithm. Together, these results suggest a need to rethink how large-scale algorithms use data on human behavior, especially in online contexts where so much of the measured behavior might be quite automatic.
Keywords: Algorithmic Bias; Behavioral Economics; Consumer Choices
JEL Codes: A12; D63; D83
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| algorithms trained on automatic behaviors (C45) | replicate and exacerbate existing biases (D91) |
| context of behavior (C92) | extent of bias (D91) |
| news feed algorithm (C45) | significant outgroup bias (C92) |
| people you may know algorithm (D85) | no detectable bias (C90) |
| automaticity of data (Y10) | bias in algorithmic outputs (C46) |
| more automatic behavior (L23) | more biased algorithms (C51) |