A discriminatory effect measured on a model's outputs, independent of any intent and of the data's apparent neutrality.
Disparate impact moves the control of discrimination from input variables to produced outputs. A model is judged on what it does, not on what it was fed, which is the only approach that survives reconstruction by proxy. Colorado legislation enshrined it by requiring quantitative testing for disparate impact even where data and model are facially neutral, a text extended in October 2025 to private passenger auto and health cover. The obstacle is perfectly ironic, testing impact by origin presupposes knowing origin, which the law often forbids collecting, so impact-based regulation remains a requirement without an instrument.
We require testing a model's disparate impact by origin while forbidding the collection of origin, asking the insurer to prove the absence of an effect it has no right to observe.
effet disparate, impact disproportionné