Signal Who Should Be Generated? Justifying Demographic Targets in Open-Ended Generation

Summary

A team led by Zeshen Zheng and colleagues published a paper on arXiv (cs.CY) on August 3, 2026, addressing what they call the 'missing-target problem' in fairness evaluation of generative AI. When a model is asked to generate something like 'a CEO in the United States' without specifying demographic attributes, existing fairness definitions assume a target distribution is already given, but the paper argues this target itself is rarely justified. The authors decompose target construction into four components: the evaluative object, prior admissibility, allocation, and operationalization. Using their benchmark AP-Bench, they find that model outputs diverge substantially from geography-derived targets, with divergence scores ranging from 0.508 to 0.606 on a 0-to-1 scale. Replacing geography-derived targets with equal-category comparators, while holding generations and measurement fixed, produces mean absolute cell-level JSD2 changes ranging from 0.279 to 0.355. The authors conclude that target construction is not a preliminary step but an integral component of fairness evaluation itself.