Note that you will always have this problem, because the data it is trained on has its own biases.