A Survey on Preserving Fairness Guarantees in Changing Environments

Real-world AI systems face a "fairness gap" between training and deployment: Most fairness research assumes training and test data come from identical distributions, but this assumption breaks down in practice. When AI systems encounter different data environments after deployment, their fairness guarantees can fail—meaning a system that appeared fair during testing may discriminate unfairly once deployed.

Distribution shift is a widespread but underaddressed problem: The paper identifies an "emergent research line" studying fairness under changing environments, suggesting this critical issue has only recently received systematic attention despite being common in real-world applications affecting human lives.

Fairness and accuracy can both degrade simultaneously under distribution shift: The survey indicates that when training and deployment environments differ, algorithmic performance suffers not only in accuracy but specifically in maintaining fairness guarantees—meaning vulnerable groups may face compounded harms.

A unified framework is needed across the field: By proposing a taxonomy of approaches for fair classification under distribution shift and connecting this work to related research areas, the survey suggests the field lacks standardized methods and benchmarks, making it difficult for practitioners to know which fairness-preserving techniques will work in their specific changing environments.

Ongoing research gaps remain in translating fairness from lab to practice: The identification of "future venues of research" indicates that while the problem is recognized, solutions for maintaining fairness guarantees across real-world distribution shifts are still developing and incomplete.