Why Could Ai Learn Harmful Biases