Researchers developed a formal framework to study bias amplification in machine learning, particularly in majority-minority learning tasks. This framework reveals how standard training can lead to stereotypical predictors that neglect minority-specific features, and provides a lower bound on the additional training required to mitigate this issue.