Skip to content

Reading a Water Map Before a Data Center Decision

For people using AI to assess a proposed data center, a water-stress map can identify an existing pressure to investigate but cannot establish the facility's water requirements or commitments.

When Lumberton, North Carolina, adopted data-center rules on November 9, residents wanted more than a general assurance about development. They asked about effects on the city’s water plant, the Lumber River and electricity rates. The Rev. Mac Legerton sought a requirement for a community-benefits agreement with any future developer. The council voted unanimously for the ordinance; the report also records residents speaking in favor. This was a vote on rules, not approval of an individual facility.

For someone using AI to prepare for such a decision, these are different assignments. Understanding pressure on a water supply is not the same as estimating a proposed facility’s demand. Understanding a proposed safeguard is not the same as obtaining a commitment to it. An assistant’s ability to explain environmental records becomes useful when those explanations connect to the particular information and obligations a decision requires.

A November 10 preprint introducing EJAI, a prototype combining environmental maps, documents and AI, offers a small study of that preparation. Its 15 nonexpert adult participants investigated hypothetical data-center proposals for a simulated public meeting. They used AI to interpret records, but still wanted facility details and community input the available material did not supply. The study did not involve Lumberton’s residents or test whether real communities secured better decisions.

What the water-stress score describes

One participant, identified as P13, wanted to know how much water or electricity the data center would use. Those figures, the participant said, would influence the recommendation. The request was not for a clearer definition of an environmental term. It was for information about the proposed facility.

The distinction is visible in baseline water stress, one of EJAI’s map indicators. The underlying Aqueduct methodology defines it as water demand relative to available renewable surface and groundwater supplies. Higher values indicate greater competition among users. The World Resources Institute, which provides the indicator, describes Aqueduct as a prioritization tool to supplement with local investigation.

A high score can make a facility’s requirements more consequential to investigate. It cannot provide those requirements. The indicator describes a surrounding water situation, while P13 needed the proposed addition to it. A project filing with a demand estimate could supply another part of the answer. Rephrasing the indicator could not.

Even the map’s geographic precision needs care. Aqueduct calculates its water-stress indicator for river sub-basins; EJAI displays county-level indicators. Neither geography automatically describes the water arrangements of an individual facility. A score on a county map is a reason to investigate, not a calculation of the project’s effect on a particular neighborhood.

AI nevertheless had an identifiable job in the study. Without it, one participant struggled with zoning acronyms in testimony and another with traffic proximity and particulate matter. With AI access, participants asked for clarification. P2 described the help as “providing context to what things mean,” while saying the assistant did not make the decision. Making a record intelligible can let someone bring it into a recommendation without settling what the recommendation should be.

Why testimony is not another map layer

The available material had several origins. EJAI’s collection held 107 documents and 44 public comments, alongside environmental indicators and tracked facilities. Of those documents, 77 were corporate reports and disclosures. The researchers did not design the collection to represent every stakeholder position. These counts describe its composition, not which source was correct or how much weight an answer gave it.

The differences mattered in participants’ reasoning. According to the researchers, P1 rejected a hypothetical proposal after combining high baseline water stress, drinking-water concerns, wastewater indicators and testimony about Indigenous cultural significance. The recommendation brought together pressure on supplies, drinking-water concerns and the significance of a place. Those considerations were not separate measurements of the same effect.

Testimony can identify a concern an environmental indicator does not record without establishing how common an experience is. A water-stress ratio, in turn, does not measure a place’s cultural importance. Preserving the contribution of each source gives someone more to reason with than treating them all as interchangeable evidence that a project is safe or harmful.

The collection also set limits on whose concerns could appear. P10 questioned whether the testimony included people with limited English and recommended waiting for their input. This was a participant’s concern about representation, not a measured finding about who had been excluded. It identifies a different kind of missing information from P13’s demand estimate. Obtaining a technical filing would not substitute for hearing those residents.

Nor did a map’s presentation guarantee scrutiny. P1 described built-in metrics as feeling official; P7 reflected on taking environmental statistics at face value. Some participants questioned corporate-sourced AI claims more readily. Explaining the corporate material carefully would leave another task open: understanding what the map itself measured and which question it could answer.

A second voice still needs a water plan

EJAI could simulate an Environmental Advocate and a Corporate Executive. These were assigned argumentative roles using the same model and retrieved evidence, not representatives of actual stakeholders. Some participants used them to consider disagreement; others found the characters simplified or committed to conclusions before examining information.

A separate technical test exposed a particularly relevant failure. Answering questions the authors wrote from their source collection, the Corporate Executive claimed that data centers use recycled water for cooling, although the retrieved sources did not support the assertion. That does not mean recycled-water cooling is impossible. It means a plausible technology had entered the answer without the evidence needed to establish its use.

The corporate role also omitted some retrieved evidence contrary to its priorities. Two-role exchanges recovered much of that coverage, but did not establish superiority over the assistant without a persona. Hearing a rebuttal can expose another consideration; it cannot create an independent observation about the facility. For the water question, the missing object remains a documented plan, not a more convincing voice describing one.

From a recommendation to an obligation

The study’s preparation briefs included requests for water-use disclosure, grid analysis, monitoring and community-benefit agreements. These were hypothetical recommendations, not negotiated safeguards. Their appearance also needs to be read alongside the assignment: the brief explicitly asked participants for missing information, trade-offs and a question for officials.

Participants tried a dashboard with maps and documents but no AI, a version with one chatbot, and EJAI’s bundle of chat, missing-information statements, simulated roles and a comparison matrix. They explored each for 8 minutes, with county and order assignments counterbalanced. The English-fluent adults were recruited from one U.S. state. The researchers report themes across conditions, identifying a condition where relevant; these accounts do not isolate the benefit of a feature or establish that AI caused conditional recommendations.

Still, the substance of those recommendations clarifies the work an assistant could support. Explaining a water-stress measure addresses understanding. Finding the proposed facility’s demand addresses a project-specific gap. Hearing missing residents addresses participation. Asking for an agreement concerns what someone would undertake to do. Progress on one does not automatically finish the others.

Legerton’s request in Lumberton gives the last task a real institutional setting, separate from the study. The question was not simply whether a developer could describe desirable practices, but whether a benefits agreement should be required. An assistant preparing someone for that discussion would need to keep a suggested cooling technology distinct from a documented water plan, and a proposed safeguard distinct from an obligation a developer has accepted.

Sources

Understand How People and Agents Work Together.

Practical essays, experiments, and lessons from building software for people and agents.

Occasional emails, sent when there is something worth reading. Free to subscribe. Unsubscribe at any time.