The reconstitution of a protected variable's effect through lawful factors correlated with it, which machine learning renders undetectable.
Indirect discrimination arises when neutral-looking variables, occupation, vehicle type, place of residence, act as stand-ins for a correlated protected characteristic. Banning the protected variable therefore does not remove its effect, a model trained on hundreds of correlates reconstitutes it unbidden, but removes the ability to measure it. This is why deliberate blinding of the model is universally judged the worst fairness strategy, though it is in substance the one the law has imposed.
A study of 2011 to 2013 data established that direct gender discrimination, once banned, reconstituted itself through occupation and vehicle type.
discrimination par substitut, proxy discrimination, variable de substitution