Learning from Misalignments