Abstract Most definitions of algorithmic bias and fairness encode decision-maker interests, such as profits, rather than the interests of disadvantaged groups (e.g. racial minorities): bias is defined as a deviation from profit maximization. Future research should instead focus on the causal effect of automated decisions on the distribution of welfare, both across and within groups. The literature emphasizes some apparent contradictions between different notions of fairness, and between fairness and profits. These contradictions vanish, however, when profits are maximized. Existing work involves conceptual slippages between statistical notions of bias and misclassification errors, economic notions of profit, and normative notions of bias and fairness. Notions of bias nonetheless carry some interest within the welfare paradigm that I advocate for, if we understand bias and discrimination as mechanisms and potential points of intervention.
Read full abstract