The reproduction of a protected variable's effect by a model that reconstructs it from its correlates, with no intent to discriminate.
Proxy discrimination is the mechanism by which a model deprived of a protected characteristic, but fed thousands of other variables, reconstructs its effect from its correlates. It requires no intent, the model's function being precisely to find any predictive correlation, and the protected characteristic is one. It renders obsolete anti-discrimination law built on lists of prohibited variables, for forbidding a variable does not forbid its effect. The only remedy is to regulate outputs rather than inputs, testing for disparate impact even on a facially neutral model, which runs into the impossibility of measuring what one forbids collecting.
A model deprived of a protected variable but fed thousands of others reconstructs the missing variable's effect from its correlates, without seeking to discriminate.
proxy discrimination, discrimination par proxy