The paper "The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning" by Sam Corbett-Davies and Sharad Goel from Stanford University critically examines the field of fair machine learning, which aims to ensure equitable outcomes in algorithmic decision-making. The authors discuss three prominent definitions of fairness: anti-classification, classification parity, and calibration. They argue that these definitions have significant statistical limitations and can sometimes harm the very groups they are intended to protect. For instance, requiring anti-classification or classification parity can lead to discriminatory decisions, while calibration, while desirable, does not guarantee equitable outcomes. Instead, the authors propose treating similarly risky individuals similarly based on statistically accurate risk estimates, which often violates anti-classification and classification parity. They emphasize the importance of addressing these limitations to advance the field of fair machine learning and highlight the need for practical solutions that align with policy objectives. The paper also discusses the challenges of constructing suitable risk estimates and the role of measurement error and sample bias in statistical risk assessment algorithms.The paper "The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning" by Sam Corbett-Davies and Sharad Goel from Stanford University critically examines the field of fair machine learning, which aims to ensure equitable outcomes in algorithmic decision-making. The authors discuss three prominent definitions of fairness: anti-classification, classification parity, and calibration. They argue that these definitions have significant statistical limitations and can sometimes harm the very groups they are intended to protect. For instance, requiring anti-classification or classification parity can lead to discriminatory decisions, while calibration, while desirable, does not guarantee equitable outcomes. Instead, the authors propose treating similarly risky individuals similarly based on statistically accurate risk estimates, which often violates anti-classification and classification parity. They emphasize the importance of addressing these limitations to advance the field of fair machine learning and highlight the need for practical solutions that align with policy objectives. The paper also discusses the challenges of constructing suitable risk estimates and the role of measurement error and sample bias in statistical risk assessment algorithms.