Fair Play for Individuals, Foul Play for Groups? Auditing Anonymization's Impact on ML Fairness

Anonymization can severely harm group fairness: The study found that anonymization techniques (k-anonymity, ℓ-diversity, and t-closeness) can degrade group fairness metrics by up to fourfold, meaning that ML systems may make systematically worse decisions for certain demographic groups after data is anonymized for privacy protection.

Privacy and fairness protections can conflict: While anonymization successfully protects individual privacy by making people harder to identify, it simultaneously undermines fair treatment across different population groups—revealing a fundamental tension between two important responsible AI goals that organizations must actively manage.

Individual fairness may appear to improve, but for the wrong reasons: Anonymization techniques showed improvements in individual fairness metrics, but this improvement stems from increased homogeneity in the data rather than genuinely fairer decision-making, suggesting that surface-level fairness improvements can be misleading.

Trade-offs require intentional decision-making: The research demonstrates that privacy, fairness, and model utility cannot all be optimized simultaneously, requiring organizations to make deliberate choices about which values to prioritize based on their specific context and stakeholder needs.

Actionable guidelines are now available: The authors provide specific recommendations for responsible AI development that account for these privacy-fairness trade-offs, with publicly available code enabling organizations to audit their own systems for similar impacts.