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

In many humanitarian and social policy settings, decision-makers must distribute limited resources—such as cash transfers or demining efforts—to those who need them most. While machine learning models are increasingly used to predict who is most vulnerable, these models cannot eliminate "aleatoric uncertainty," which is the inherent, irreducible randomness in individual outcomes. Even a perfect model will misallocate some resources because it only predicts vulnerability in terms of probability. This paper explores how to combine AI-driven predictions with traditional, direct screening (such as physical verification) to create a more efficient two-stage allocation system.

The Two-Stage Allocation Framework

The authors propose a framework that balances the speed of AI with the precision of physical verification. In the first stage, a decision-maker uses a limited "screening budget" to verify the true status of a specific subset of the population. In the second stage, the resource is allocated: screened individuals who are confirmed to be vulnerable receive the resource, and the remaining budget is distributed to the highest-risk individuals identified by the AI model. By observing the true outcomes of the screened group, the decision-maker resolves the inherent uncertainty for those individuals, ensuring that resources are not wasted on those who do not actually need them.

Where to Screen for Maximum Impact

A key finding of the research is that the most effective screening strategy does not target the highest-risk or lowest-risk individuals. Instead, the optimal approach is to screen individuals who fall into an "uncertainty band" at the margin of the AI’s allocation threshold.
Screening the very highest-risk individuals is inefficient because the AI is already confident they should receive the resource. Conversely, screening the lowest-risk individuals is wasteful because it is unlikely to uncover a vulnerable person. By focusing on the "middle" group—those whose risk scores are close to the cutoff for receiving aid—the decision-maker can use their limited screening budget to clarify the most ambiguous cases, thereby maximizing the overall impact of the resource distribution.

The Value of Screening

The study establishes that the efficiency gains from screening are directly linked to the level of aleatoric uncertainty in the population. When a population has high irreducible uncertainty, the AI model is less reliable, making the marginal value of physical screening significantly higher.
The authors also identify a "law of diminishing returns" regarding screening: the efficiency gains from adding more screening capacity are concave, meaning that while initial investments in screening provide significant improvements, these benefits eventually level off. This provides a formal way for policymakers to determine how much they should invest in physical verification versus relying on algorithmic predictions, depending on their specific budget constraints and the nature of the population they are serving.

Practical Applications

To demonstrate the effectiveness of this framework, the authors applied it to real-world scenarios, including income-based social protection programs and humanitarian demining operations in Colombia. In these contexts, the tension between the cost of verification and the need for accurate targeting is a major operational challenge. The research provides a concrete, code-supported tool that allows practitioners to input their own risk scores and budget constraints to calculate the optimal screening and allocation policy, helping them navigate the inherent limits of AI-driven decision-making.