Proxy Variable

A proxy variable is a feature or data point that, while not directly measuring a protected characteristic, is strongly correlated with one. Common proxy variables in AI hiring include zip code (proxy for race), name (proxy for ethnicity/gender), graduation year (proxy for age), university attended (proxy for socioeconomic status), and gaps in employment history (proxy for gender/disability). AI models can learn to use proxy variables to make predictions that effectively discriminate based on protected characteristics, even when those characteristics are not included as direct inputs. Identifying and mitigating proxy variables is a critical part of AI bias auditing.