Disparate Impact
The disproportionately harmful effect a facially neutral practice has on a protected class. Insurance regulators address the same problem under other names.
Disparate impact occurs when a policy or practice that seems neutral on its face produces disproportionately negative outcomes for members of a protected class. In insurance, this might show up as higher denial rates for one racial group, higher premiums for residents of certain neighborhoods, or worse claims outcomes for one gender.
The concept is central to AI fairness testing. Even if no protected characteristic is an explicit input, a model can produce disparate impact through proxy variables like geography, credit data, or occupation.
The term itself comes from employment and housing law, and it is worth knowing that the insurance regulators borrowed the structure without the name. The NAIC Model Bulletin never uses the phrase; it asks for bias analysis and for methods that detect unfair discrimination. The NAIC AI Systems Evaluation Tool never uses it either, or “proxy.” NYDFS Circular Letter No. 7 says “disproportionate adverse effect” throughout, and the only place anything like “disparately impacted” appears in it is a footnote quoting the American Academy of Actuaries. When describing your own testing to an examiner, their words travel better than this one.
If a proxy variable has disproportionate adverse effects, dropping it is not the only response, as we read the letter. Its proxy assessment, a proxy test for external consumer data, asks whether legitimate business necessity requires the use. Its comprehensive assessment, also covering AI systems, ends with a documented search for less-discriminatory alternatives; if one exists, the use should be modified and retested. See our guide to the proxy test.