Saarland University's First-See-Then-Design Rewrites How AI Fairness Gets Measured
A new fairness framework rejects prediction-based parity metrics in favor of stakeholder utilities, showing stochastic policies expand Pareto trade-offs.
Takeaways
- New multi-stakeholder fairness framework shifts evaluation from predictive metrics to realized welfare across DM, DS, and social planner.
- Fairness gets defined via distributive justice principles like Egalitarian and Rawlsian, not demographic parity.
- Post-hoc multi-objective optimization exposes the full Pareto front before committing to a policy.
- Stochastic policies provably dominate deterministic thresholds when DM utilities are asymmetric or misaligned with groups.
- Experiments on five datasets show large hypervolume gains, especially for group-specific stochastic policies.
- Paper accepted at FAccT '26; implementation released in official repository.
Most fairness work in machine learning treats the problem as a tug-of-war between accuracy and a statistical parity metric like demographic parity or equality of opportunity. A new paper accepted at FAccT '26 argues that this framing measures the wrong thing entirely, and that fairness should be evaluated on the actual welfare decisions produce, rather than on the properties of the predictions that generate them.
The authors, from Saarland University and Zurich University of Applied Sciences, propose a multi-stakeholder framework called First-See-Then-Design, which explicitly models the utility of decision-makers, decision subjects, and a social planner, then exposes the full set of Pareto-optimal policies before anyone commits to one. Along the way they prove, and empirically demonstrate, that simple stochastic decision rules can strictly dominate deterministic thresholds under certain utility structures.
Where predictive metrics quietly fail
The standard pipeline runs prediction, then decision, then outcome, then welfare. By collapsing this chain into predictive metrics, prediction-centric approaches implicitly assume that fair predictions translate into fair outcomes, an assumption that does not generally hold. Equalizing true positive rates across groups says nothing about whether two groups actually experience comparable benefit from a favorable decision, since costs of rejection and gains from approval can differ substantially across demographics.
The paper shows this concretely on the German Credit dataset. When the authors take policies that are Pareto-optimal in the predictive space (accuracy vs. equality of opportunity) and project them into utility space, they land in a small suboptimal region. Directly optimizing utilities yields substantially richer Pareto fronts, expanding the set of viable policies available at deployment.