Teaching AI to read the Clean Water Act

One of the big challenges in permitting reform is that it’s really hard to forecast how changing one piece of the massive US legal/regulatory permitting system will play out. At a more micro-level, it’s also a headache for developers to try to figure out what environmental requirements their project actually triggers. Specifically, when contemplating a new project, developers do not know if the Clean Water Act applies until the Army Corps of Engineers completes an assessment. In a new working paper, Greenhill, Walker, and Shapiro train a deep learning model to evaluate whether Clean Water Act provisions apply to a given parcel, using Army Corps of Engineers determinations tunder several recent rulemakings. They find their model is 65 times better at identifying regulated sites than the leading geophysical approach. This helps in analyzing past regulatory shifts (the authors find Sackett deregulated roughly a third of all previously regulated waters), helps in generating high-quality projections of proposed regulations before they’re implemented, and could potentially help developers better predict if they’re actually subject to CWA restrictions.