AI Fairness Beyond Complete Demographics: Current Achievements and Future Directions

Complete demographic data is often unavailable in practice: The paper highlights that most fairness methods assume access to full demographic information, but legal constraints (like privacy regulations) and discrimination risks make collecting complete demographic data impractical in real-world AI systems, creating a significant gap between research and deployment.

A new framework is needed for fairness without complete demographics: The authors introduce a novel taxonomy of fairness concepts specifically designed for scenarios with incomplete demographic information, clarifying how different fairness approaches relate to and differ from one another when demographic data is limited or missing.

Current bias mitigation techniques have fundamental limitations: Existing methods that rely on complete demographic information cannot be directly applied to real-world systems, meaning many proposed fairness solutions may not actually protect against discrimination in deployed AI systems that lack comprehensive demographic data.

Privacy and fairness create competing tensions: The paper identifies a key challenge: collecting complete demographic information to improve fairness can itself reinforce discrimination and violate privacy principles, requiring AI developers to balance these competing ethical concerns rather than simply maximizing demographic data collection.

Significant research gaps remain: The survey identifies open questions about how to achieve fairness in AI systems with incomplete demographics, suggesting this is an emerging area where more practical solutions are urgently needed for responsible AI deployment.