Proxy Discrimination
Discrimination that happens when a neutral variable, such as zip code or credit data, correlates with a protected class and produces unfair outcomes.
Proxy discrimination occurs when a model uses a variable that appears neutral but is correlated with a protected class, producing unfair outcomes for that group. Common examples include zip code, credit score, occupation, and certain geographic or lifestyle variables.
The NYDFS proxy test and the NAIC Model Bulletin both require insurers to screen for these correlations. The test is not whether a variable is a perfect proxy, but whether it is correlated enough with a protected class to produce materially different outcomes. If it is, the carrier must show the variable is required by a legitimate business need and that no less-discriminatory alternative exists.
Proxy discrimination is one of the hardest problems in insurance AI because the most predictive variables often turn out to be proxies for protected characteristics. See our guide to the proxy test and glossary entries on protected class and disparate impact.