The Limits of AI-Driven Allocation: Optimal Screening under Aleatoric Uncertainty

Researchers present a framework for optimally combining algorithmic risk scoring with direct verification screening in resource allocation decisions. The study demonstrates that even perfect predictive models cannot eliminate misallocation due to irreducible uncertainty about individual vulnerability, and shows that screening is most effective when focused on borderline cases rather than high-risk units.

Machine learning has transformed humanitarian and policy resource allocation by replacing costly manual verification with algorithmic risk scoring. However, this research reveals a fundamental constraint: aleatoric uncertainty—inherent randomness in individual outcomes—creates an irreducible floor on misallocation regardless of model accuracy. This distinction matters because it reframes the allocation problem from purely improving predictions to strategically combining algorithmic and direct verification methods.

The two-stage framework addresses a practical tension in resource allocation. Screening observes true outcomes for selected units before final distribution, enabling targeted correction of algorithmic decisions. The research reveals that optimal screening focuses on units near algorithmic decision boundaries—those where risk scores provide least confidence—rather than highest-risk populations. This counterintuitive finding reflects the diminishing returns to additional information when risk scores already strongly indicate vulnerability.

The complementarity or substitution between screening and algorithmic methods depends critically on population-level uncertainty. In heterogeneous populations where vulnerability is difficult to predict, screening becomes increasingly valuable. Conversely, in homogeneous populations where algorithms perform well, screening provides marginal benefit. This characterization has immediate operational implications for humanitarian organizations and social protection programs facing budget constraints.

The Colombia demining and income-protection case studies ground abstract findings in real policy contexts where screening costs and allocation efficiency directly compete for limited budgets. Organizations can now assess whether their populations exhibit sufficient aleatoric uncertainty to justify verification investments. This framework enables more rigorous cost-benefit analysis in humanitarian contexts while establishing research standards for evaluating algorithmic allocation systems beyond prediction accuracy alone.

• →Perfect algorithmic predictions cannot eliminate misallocation because individual vulnerability outcomes contain irreducible uncertainty independent of predictor quality.
• →Optimal screening targets units at the algorithmic decision margin rather than highest-risk populations, maximizing information gain per verification cost.
• →Screening and algorithmic targeting act as complements when population uncertainty is high, suggesting their relative value depends on specific operational contexts.
• →The framework enables quantitative assessment of screening investments in humanitarian programs by measuring population-level aleatoric uncertainty.
• →Two-stage allocation combining algorithms with targeted verification outperforms pure algorithmic or pure screening approaches across tested scenarios.