We characterize the problem of achieving group sufficiency under label bias, and introduce a regularizer that restores fairness without sacrificing accuracy.