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

Gender prediction can be technically valid (producing accurate measurements) while still being ethically illegitimate (causing harm), and the paper argues these are separate concerns that researchers must evaluate independently rather than assuming one determines the other.

The harms of gender prediction differ by target group: the research distinguishes between sexism targeting women and femininity (where gender imputation may be justified for fairness research) versus sexism targeting transgender and nonbinary people (where it causes direct harm and should be avoided).

Gender imputation for fairness research requires strict justification: practitioners should only use algorithmic gender prediction when it would achieve anti-discrimination benefits that "cannot be achieved through other reasonable means," with harms minimized as much as possible—not as a default practice.

Real-world applications reveal the stakes: the paper examines three concrete cases (auditing bias in image generation, measuring film disparities, and inferring gender from names) to show how gender imputation decisions have different legitimacy implications depending on context and who is affected.

Responsible AI requires distinguishing between technical soundness and ethical appropriateness: a measurement can be statistically valid while remaining harmful, and researchers must actively evaluate legitimacy alongside validity rather than treating technical accuracy as sufficient justification for using predicted gender labels.