Artificial Intelligence in Environmental Protection: The Importance of Organizational Context from a Field Study in Wisconsin

TL;DR

This study investigates how organizational context shapes the deployment of AI for environmental protection by conducting parallel field trials in Wisconsin with WDNR (regulator) and ELPC (advocacy group) using a satellite-imagery–based land-application detector. The approach demonstrates substantial efficiency gains in detecting manure spreading and reveals that many detections fall outside existing regulatory scopes, highlighting policy design gaps. A key contribution is the demonstration of a regulatory Rashomon effect, where identical AI outputs are interpreted differently by organizations with distinct goals, underscoring the need for governance and alignment between technology and environmental policy. The findings emphasize the value of human-in-the-loop collaboration to improve precision and policy-relevant insight, while also calling for policy reforms to better match environmental risk with regulatory thresholds.

Abstract

Advances in Artificial Intelligence (AI) have generated widespread enthusiasm for the potential of AI to support our understanding and protection of the environment. As such tools move from basic research to more consequential settings, such as regulatory enforcement, the human context of how AI is utilized, interpreted, and deployed becomes increasingly critical. Yet little work has systematically examined the role of such organizational goals and incentives in deploying AI systems. We report results from a unique case study of a satellite imagery-based AI tool to detect dumping of agricultural waste, with concurrent field trials with the Wisconsin Department of Natural Resources (WDNR) and a non-governmental environmental interest group in which the tool was utilized for field investigations when dumping was presumptively illegal in February-March 2023. Our results are threefold: First, both organizations confirmed a similar level of ground-truth accuracy for the model's detections. Second, they differed, however, in their overall assessment of its usefulness, as WDNR was interested in clear violations of existing law, while the interest group sought to document environmental risk beyond the scope of existing regulation. Dumping by an unpermitted entity or just before February 1, for instance, were deemed irrelevant by WDNR. Third, while AI tools promise to prioritize allocation of environmental protection resources, they may expose important gaps of existing law.

Figures (11)

• Figure 1: Example of satellite imagery (left) and ground-verified (right) detection of manure spreading in Grant County, WI.
• Figure 2: Process metrics from the ELPC trial. (A) Follow-up rate by model score. (B) Among detections visited, ELPC verifiers were able to see the location from public roads at similar rates across model scores. (C) Distribution of days to follow-up.
• Figure 3: Detection validation rates by model confidence for WDNR (A, B) and ELPC (C). (A) and (C) show the overall confirmation rate among all detections sent to each organization while (B) shows the confirmation rate only among detections sent to WDNR that passed initial desk screening and were investigated by specialists, illustrating the value of expert review of model outputs.
• Figure 4: Determination of regulatory compliance for the 64 spreading events confirmed by WDNR. Only a small fraction (17%) were found to be non-compliant CAFO spreading.
• Figure 5: Example of satellite imagery confirming a manure application detected by our model as of February 2 had been spread in late January, prior to the general restrictions.