I think what he's suggesting is feeding the input data that differs only on that variable, and if the algorithm produces differing output, then it is by definition racist (and then you send the authors back to the drawing board to figure out why).