Joint Optimization of AI Fairness and Utility: A Human-Centered Approach

Fairness tradeoffs are unavoidable: The paper demonstrates that different fairness criteria "sometimes cannot be simultaneously satisfied," meaning developers must make deliberate choices about which fairness goals to prioritize rather than assuming all can be achieved together.

Human policymakers should guide AI decisions: Rather than letting technical experts alone decide fairness-accuracy tradeoffs, the framework emphasizes acquiring input from human policy makers on how to balance competing objectives, recognizing that these are ultimately value-based decisions that should reflect societal priorities.

Fairness often comes at a cost: The research acknowledges that "achieving fairness often requires sacrificing other objectives such as model accuracy," highlighting that organizations cannot simply add fairness without accepting potential performance tradeoffs that deserve transparent discussion.

Current bias mitigation methods lack stakeholder input: The paper notes that while the AI research community has "proposed many methods to measure and mitigate unwanted biases, few of them involve inputs from human policy makers," suggesting a gap between technical solutions and real-world decision-making contexts.

Preference elicitation is essential: The framework proposes specific methods for "eliciting preferences" from policymakers and then optimizing AI systems accordingly, treating human values as a critical input rather than an afterthought in AI development.