Amazon Just Settled a Pollution Case AI Reports Can’t Measure

Last month, Amazon agreed to pay $20.5 million to settle a class action brought by residents of Morrow County, Oregon, who say they cannot drink the water from their wells. The case, Pearson v. Port of Morrow, was filed in federal court in February 2024. The plaintiffs are among roughly 45,000 people in the Lower Umatilla Basin who depend on groundwater. The nitrate levels in that groundwater have been linked to cancer, thyroid disease, and harm to infants. Many of the residents are low-income and Latino. Amazon denied the allegations and admitted no guilt.

Amazon has run cloud computing data centers at the Port of Morrow since 2011, according to Rolling Stone. Those data centers use water to cool servers running around the clock. Reporters found that wastewater from the facilities entered the Port’s broader wastewater system. The Port then sent that nitrate-heavy effluent to nearby farmland, which has contributed to years of groundwater contamination. A 2025 Oregon Capital Chronicle investigation found the wastewater had been overapplied in the winter for years, letting excess nitrate seep into the aquifer. Those servers run the same workloads that AWS markets on a page called “AI Factories”. Nothing in Amazon’s public reporting says how much of that water, or that harm, is attributable to AI.

That is a governance problem, and it is a wider one than Amazon. A study published in Patterns in December 2025 found that data center operators do not separate AI workloads from non-AI workloads in their environmental disclosures. The researcher, Alex de Vries-Gao of VU Amsterdam’s Institute for Environmental Studies, concluded that fixing this requires new policies to compel additional disclosure. The reporting categories operators use were built before AI was its own category, and they have not been updated to treat it as one. Any estimate of AI’s carbon or water footprint has to be approximated from whole-facility metrics. Until then, the environmental cost of running AI at any specific data center is structurally hard to see in the public record. Every responsible-AI disclosure from a major AI company currently rests on that record. When Anthropic, OpenAI, or Mistral describe their environmental commitments, they cite their cloud providers’ efficiency metrics, and those metrics are averages across every workload a cloud provider runs. The best numbers these companies have were not built to answer the more specific question of what any single AI workload costs. An AI company can publish a climate commitment that is both signed in good faith and impossible to verify against what is actually happening at the facility level.

I reported on a version of this structural problem at the Port of Los Angeles, where half the port’s air quality monitoring sensors were down during the worst period of pandemic-era pollution, and the port’s response was that it had enough data from other sources. Andrea Hricko, a retired USC professor I interviewed, put it plainly. The instruments were not working, so it looked like residents were breathing clean air. The monitoring regime produced the appearance of oversight without the substance of it. Morrow County is the same failure at a different scale. The reporting instruments were never calibrated to measure what AI is costing the water table in Eastern Oregon.

The same day the settlement was announced, Morrow County’s Planning Commission approved a new Amazon data center campus east of the Boardman airport. According to the Tri-Cities Area Journal of Business, the project, called PDX 154, will span 816,000 square feet across four buildings with two wastewater evaporation ponds. Planners attached conditions requiring Amazon to address stormwater and wastewater impacts. The metrics used at the existing facilities could not show the harm until a federal lawsuit forced a settlement. The new conditions will be measured against the same metrics. The build-out is running years ahead of the accounting.

A settlement without an admission is not an accountability event in the governance sense. It closes the litigation and transfers money. It does not establish that the reporting regime worked, because the reporting regime is not what produced the outcome. What produced the outcome was Oregon residents who could not drink their water hiring lawyers and spending two years in federal court. If that is the mechanism by which AI’s environmental cost becomes visible, then the sustainability reports are not governance. They are corporate disclosures that can coexist with serious harm because the only thing that tests them is litigation, and litigation sits outside the reporting mechanisms.