Policy Context

Many low- and middle-income countries rely heavily on tax revenues collected at their borders, making corruption at customs a costly issue that hinders development. Can new AI tools strengthen state capacity and increase tax revenue? A team of researchers has developed a machine learning algorithm that predicts import tax evasion at shipping ports in Paraguay, where according to their analysis, over $200 million in additional tax revenue could be collected if misreporting was curtailed. The team’s new algorithm is over three times more accurate than current methods in flagging shipments with customs fraud. It does this with 70% fewer inspections, greatly reducing both opportunities for corruption by officials and inspection delays.

Study Design

The team of researchers (comprised of CEGA-affiliates Fred Finan and Ernesto Dal Bó, Laura Schechter at UW-Madison, and Raul Duarté at Harvard) will run a randomized controlled trial to determine the inspection algorithm’s causal impact at ports in Paraguay. Each day they will take the set of shipments flagged by the algorithm and assign a random subset of them to be inspected. They will also take a random subset of shipments that are not flagged by the algorithm and assign those to be inspected. This RCT will be run at two sets of ports. In the first, both port inspectors and a reverification team will inspect the chosen shipments. In the second, only the reverification team will inspect the chosen shipments. The reverification team acts as a ground truth and allows the measurement of both the algorithm’s accuracy and port inspectors’ honesty and capacity.

Results and Policy Lessons

Results are forthcoming.