you don't need any evidence as it is obvious that ethnicity is part of the feature vector being fed into the system. Even if we imagine unimaginable - that it

you don't need any evidence as it is obvious that ethnicity is part of the feature vector being fed into the system. Even if we imagine unimaginable - that it (together with other highly correlating data like place of birth) is not fed into the system from the start, the face recognition and other input feature vectors generating modules would produce the configuration in the parameter space where basic generic clustering at the pre-processing stage would also naturally cluster along ethnicity lines thus still making ethnicity an input for higher levels of the system. How this input will play inside that giant neural/statistical net wouldn't be possible to say as debugging such nets is an endeavor in itself. Previous implementations of such a pre-crime detection neural nets - using homo sapiens brain as hardware - produced "being black is a probable cause", and nobody was able to debug it in sufficient detail. Would we be able to produce better results using silicon based hardware?