>It always picks its own

Makes sense to me, in that its own output would align closer to its own training set