Signal Scarcity and Predictive Uncertainty: Implications for Societal Resource Allocation

Summary

Shafkat Farabi, Patrick J. Fowler, and Sanmay Das published a paper on arXiv (cs.CY) on August 4, 2026, examining what happens to scarce societal resource allocation when predictive uncertainty differs systematically across a population, such as when machine learning models have significantly different accuracies across demographics. The authors formulate a novel mathematical model of scarce resource allocation accounting for heterogeneous predictive uncertainties and analyze implications for allocation mechanisms and realized population-level benefits. They find that when resources are very scarce, maximum marginal benefit (MMB) prioritization favors individuals with lower predictive uncertainty even at identical underlying initial states, but this prioritization flips when resources are abundant, targeting higher-uncertainty individuals instead. The paper illustrates these implications using the PISA educational testing dataset, with relevance to public education resource allocation, medical triage, and homelessness services under the theory of local justice. The authors raise a new ethical dilemma — whether it is just to allocate a resource to one person over another based solely on predictive uncertainty about their futures — and assess efficiency losses under both MMB and vulnerability-first prioritization, finding losses across all resource levels but particularly in low-resource settings.