Increasing Revenue Collection with Computer Vision: Experiments in Pakistan

Economic growth in developing countries is often limited by the state's inability to raise tax revenue, and property tax collection is particularly important given its revenue is typically used to finance critical public goods and services. However, in many developing countries, tax administration systems rely on infrequently updated and out-of-date property tax valuations, and tax officials often employ significant discretion when assessing properties. These factors can lead to errors that could increase tax leakages or lower citizen trust in the state. This study addresses this challenge in two steps: first, by developing a computer vision algorithm that can use property images to predict property assessments and second, by testing how well the algorithm performs in identifying properties for reassessment in a “human vs. machine” randomized controlled trial. In the trial, tax officials will be randomly assigned to either a “status-quo” group, where they will be asked to choose which properties to reassess based on only their local knowledge, or a treatment group, where they will use results from the algorithm to identify a similar number of properties for re-assessment. We will measure the impact of the intervention on outcomes such as the misclassification rate and total revenue collected.