Deep Learning Study Forecasts How Sackett Ruling Shrinks U.S. Wetland Protection

Researchers from UC Berkeley, Yale, Stanford’s Global Policy Laboratory, and NBER developed a deep learning method to predict how environmental regulations will affect wetlands and streams before they are implemented. Their analysis shows the Supreme Court’s Sackett ruling significantly reduces Clean Water Act protections, removing federal oversight from millions of acres of wetlands and hundreds of thousands of stream miles.

Researchers from the University of California, Berkeley, Yale University, the Global Policy Laboratory at the Stanford Doerr School of Sustainability, and the National Bureau of Economic Research (NBER) have developed a new way to predict how environmental regulations could affect rivers and wetlands before the rules are actually implemented. Their study uses deep learning to estimate how changes to the United States Clean Water Act could reshape federal protection of waterways. The approach aims to solve a long-standing challenge in policy analysis: forecasting the real impact of laws that have not yet taken effect.

For decades, policymakers have relied on expert judgment and environmental maps to estimate how regulations might affect wetlands, streams, and rivers. But these methods are often uncertain. Maps of wetlands and hydrological systems were not designed to determine legal jurisdiction, and they frequently miss smaller water bodies or contain outdated information. As a result, predictions about regulatory coverage can vary widely, creating confusion for developers, environmental groups, and government agencies.

The new research suggests that artificial intelligence could help close this gap by analyzing patterns in past regulatory decisions and using them to predict future outcomes.

Why Predicting Environmental Laws Is So Difficult

Environmental regulations often have huge economic and ecological consequences, but predicting their impact is complicated. Many policies are debated and evaluated before any real-world data exist about how they will work. Traditionally, analysts have relied on geophysical models that estimate which wetlands or streams might be protected based on environmental characteristics such as water flow, wetness, or connectivity to larger rivers.

However, these models do not always reflect how regulators actually interpret the law in practice. Federal agencies evaluate specific locations and determine whether they fall under the Clean Water Act. These decisions can depend on many factors beyond simple environmental features.

The study points out that this mismatch between environmental maps and legal decisions often leads to inaccurate predictions about how regulations will apply across the country.

How Artificial Intelligence Can Predict Regulations

The researchers used deep learning models trained on thousands of past federal regulatory decisions known as Approved Jurisdictional Determinations. These decisions, made by the U.S. Army Corps of Engineers, determine whether a particular site falls under federal water pollution rules.

The key innovation in the study is a method called relabeling. Because new regulations have not yet produced real decisions, the researchers modify past regulatory outcomes to simulate how those same sites would be judged under a new rule. The deep learning system then studies these modified outcomes to learn how regulatory decisions might change.

This allows the model to generate “ex ante” predictions, meaning forecasts made before a policy is implemented. The researchers also built a second model trained on real decisions made after policy changes, allowing them to compare predicted outcomes with actual results.

The results show that deep learning models are far more accurate than traditional geophysical approaches when predicting regulatory jurisdiction.

What the Sackett Ruling Means for U.S. Waters

The study applied its models to analyze the impact of the Supreme Court’s Sackett v. EPA decision, which significantly narrowed federal authority over wetlands. The ruling requires that wetlands have a clear surface connection to navigable waters to qualify for protection under the Clean Water Act.

Using deep learning projections, the researchers estimate that the ruling substantially reduces federal protection for waterways. Roughly one quarter of stream miles and slightly more than a quarter of wetland areas remain under federal jurisdiction. Compared with earlier interpretations of the law, about one-third of previously protected waters are no longer regulated.

This change affects hundreds of thousands of miles of streams and millions of acres of wetlands across the United States. The largest reductions occur in floodplains, isolated wetlands, and seasonal streams that flow only during rainfall or snowmelt.

The impact varies across regions. Coastal wetlands and areas with dense wetland networks see large declines in protection, while major rivers and lakes generally remain regulated.

What This Could Mean for Policy and Development

The findings highlight the growing role artificial intelligence could play in regulatory analysis. More accurate predictions of regulatory scope could help policymakers understand the environmental consequences of legal changes before they occur.

Developers and companies could also benefit from clearer forecasts. Businesses often need to know whether their projects require federal permits under the Clean Water Act. Uncertainty about regulatory boundaries can delay construction projects and increase costs.

At the same time, environmental groups could use predictive tools to identify ecosystems that might lose protection under proposed policy changes.

The researchers believe that the same approach could be used to analyze many other regulations, from environmental protection laws to financial and labor policies. By combining historical regulatory decisions with machine learning, governments may gain a clearer understanding of how future laws will shape landscapes and economies before those laws come into force.