Signal Algorithmic Gender Prediction Is Illegitimate, But Gender Imputation Can Yield Valid Measurements

Summary

AI ethics researchers have criticized the algorithmic prediction of an individual's gender as illegitimate in itself. Other researchers, however, actually rely on such predicted gender labels to study gender disparities and to build algorithmic fairness techniques. This paper tries to resolve that clash by distinguishing two separate ways gender prediction can go wrong. One is being illegitimate as an individual-level intervention, and the other is being invalid as a population-level measurement. The authors argue that this distinction lets researchers use gender imputation for aggregate fairness analysis without thereby endorsing individual-level algorithmic gender assignment. The paper is a methodological and ethics contribution rather than an analysis of a specific real-world deployment.