You would have to show that gender/race were not inputs to the machine learning algorithm. Correlations with inputs that were not restricted would not be a problem (in many situations ZIP code is "close enough" for race.)