Preserving Fairness in AI under Domain Shift