The Robot Made Me Do It: An Experiment on Inspector Adoption of ML Model Predictions for Clean Air Act Enforcement

Facility inspections are central to environmental regulation but are costly, infrequent, and detect only a small share of violations. Since 2015, the researchers have partnered with the U.S. Environmental Protection Agency (EPA) to apply machine learning (ML) to inspection targeting. A prior ML-based intervention increased detection of hazardous waste violations by 82 percent and was scaled nationwide. At EPA’s request, the researchers created an ML model for Clean Air Act (CAA) enforcement, which the researchers estimate can almost double violation detection rates. The researchers propose implementing a randomized controlled trial by providing staged access to ML-based risk predictions to inspectors as a tool to support local CAA inspection processes. To measure the impact on inspection and environmental outcomes, the researchers will use administrative and satellite data. The researchers will also study the human angle of AI adoption by conducting baseline and endline inspector surveys and interviews to document changes in inspector beliefs and inspection strategies, potential adoption barriers, and limitations. The study will provide insight to regulators seeking cost-effective enforcement strategies and shed light on the integration of AI by individual employees into high-stakes public-sector work. Gains in environmental enforcement would disproportionately benefit communities historically overexposed to air pollution, with important implications for environmental justice.