02 What The Evidence Shows Complete transcript and visual descriptions Slide 1: What Three Peer-Reviewed Studies Show—and Do Not Show Visual description: Title slide introducing three peer-reviewed studies with different evidence roles. We now turn to three recent peer-reviewed papers. They do not all answer the same question. Xiao and colleagues model environmental impacts of AI servers in the United States. Lei and colleagues assess why workload-level water estimates vary. Jager and Yoon screen possible water-resource and biodiversity risks from data centers. Slide 2: Different Evidence Types Answer Different Questions Visual description: Three columns compare a modeled projection, a cross-workload assessment, and a risk screen. Each evidence type answers a different question and has a different limitation. Professor Wen's water-scientist lens begins by naming the evidence type. A projection is not a meter record. A cross-workload comparison is not one commercial service. A screening-level risk is not a measured local impact. We can learn from each paper without treating their evidence as interchangeable. Slide 3: Xiao et al.: A Modeled U.S. AI Projection Visual description: Modeled U.S. scenarios estimate 731 to 1,125 million cubic meters per year, with about 71 percent indirect and 29 percent direct water use. Xiao and colleagues published an analysis in *Nature Sustainability* in 2025. In their modeled U.S. scenarios, the estimated annual water footprint of AI servers ranges from **731 to 1,125 million cubic meters**. The pathway split is especially useful. In the modeled result, approximately **71 percent is indirect**, associated with electricity generation, and **29 percent is direct**, associated with the data-center site. This challenges the idea that only visible on-site cooling matters. But keep the qualifier attached. These are modeled national projections, not meter readings from every U.S. facility. The proportions should not be applied automatically to a proposed Central New York data center. They also should not be converted into one universal “water per prompt” value. Facility design, computing demand, grid mix, location, and scenario assumptions affect the estimate. What does this paper support? A claim that electricity-related water can be a major part of an AI-server footprint in a U.S. projection. What does it not support? A claim about exact local use or demonstrated local water-quality harm. Slide 4: Lei et al.: Why One Number Cannot Fit Every Workload Visual description: Water estimates vary by more than 10,000-fold across assessed workload conditions, linked to more than 1,000-fold variation in water per electricity and about 10-fold efficiency variation. Lei and colleagues published their assessment in *Resources, Conservation and Recycling* in 2025. They examine how workload efficiency, server utilization, cooling, climate, location, and electricity supply shape water estimates. Across the conditions assessed, workload-level water use varies by more than **10,000-fold**. The authors connect this wide range to more than **1,000-fold variation in water use per unit of electricity** and about **10-fold variation in workload efficiency**. This does not mean that one ordinary prompt has a simple, directly measured 10,000-fold swing everywhere. It means the result depends strongly on the conditions and system boundary being compared. A short task on efficient equipment, in a facility using one cooling design and one electricity mix, cannot be assumed equivalent to a different task, facility, climate, or grid. The practical lesson is not “numbers are useless.” It is “numbers need context.” Before comparing two estimates, ask whether they count the same activities, use the same definition of water, and represent similar workloads and locations. Slide 5: Jager and Yoon: Screen Potential Water and Ecosystem Risks Visual description: A screening framework identifies potential pathways including water stress, altered flows, warm or lower-quality discharge, intake effects, and habitat concerns. Jager and Yoon's 2026 paper in *Water Biology and Security* broadens the view from a water-footprint total to water resources and biodiversity. Their screening-level framework identifies pathways that deserve closer attention. These include water stress and altered flows, warm or lower-quality discharge, and risks to aquatic organisms connected with water intake and habitat conditions. This paper matters because water quantity alone cannot answer a water-quality question. A withdrawal total does not tell us the temperature or chemistry of return flow. It also does not describe stream flow, habitat, or species sensitivity. At the same time, screening is an early step. It identifies where a risk could be possible and where closer study may be useful. It does not demonstrate that the fictional facility has harmed drinking water or fish. The paper also covers data centers broadly, including AI and other computing demands, rather than isolating one AI service. A calibrated statement is: “The paper identifies potential water-quality and ecosystem pathways that a local review should investigate.” An uncalibrated statement is: “The paper proves this proposed facility will damage the stream.” Slide 6: The Same Word 'Water' Can Hide Different Evidence Visual description: A national projection is not a local meter record; a quantity estimate is not a water-quality sample; and a risk pathway is not an observed ecological impact. Place the three studies side by side. Xiao and colleagues help estimate direct and indirect pathways in modeled U.S. AI scenarios. Lei and colleagues explain why workload estimates vary across technical and geographic conditions. Jager and Yoon identify potential quantity, quality, and ecosystem risk pathways. Now notice the non-equivalences. A **national projection is not a local meter record**. A **quantity estimate is not a water-quality sample**. A **potential risk pathway is not an observed ecological impact**. This is where claim-checking supports environmental reasoning. When an AI-generated explanation gives you a confident sentence, ask three short questions. What kind of evidence is it using? Does the wording preserve the study's system boundary? What evidence would be needed to move from a broad possibility to a local conclusion? Slide 7: Professor Wen's Lens: What Is Missing Locally? Visual description: Six local evidence needs lead to a local decision: cooling, source and volume, electricity mix, blowdown chemistry and temperature, permits, and receiving-water conditions. Return to the fictional Central New York proposal. The papers frame questions, but the decision requires local inputs. Ask for the **cooling technology** and operating schedule. Ask for the **source water** and seasonal withdrawal and consumption volumes. Ask about the **electricity mix**, because indirect water may occur outside the facility's watershed. For water quality, ask about **blowdown chemistry**, discharge temperature, treatment, and the discharge route. Review **permit limits** and compliance monitoring. Finally, examine **receiving-water conditions**: stream flow, temperature, existing stresses, habitat, sensitive organisms, and other users. The missing evidence should match the claim. To evaluate a scarcity claim, request seasonal quantity and source information. To evaluate a drinking-water claim, request chemistry, treatment, and downstream intake information. To evaluate a fish-impact claim, request temperature, flow, habitat, and biological evidence. We should not promise that a facility is harmless before seeing this evidence. We also should not declare damage as a fact before measuring it. The responsible position can be: “We cannot yet determine the local impact; here is the evidence needed next.” Slide 8: Your Task: Calibrate a Community Recommendation Visual description: Four requirements for a 50-to-75-word community recommendation. You will now evaluate three claims using the evidence cards, then advise the planning board in 50 to 75 words. A four-bullet response is equally acceptable. Cite at least two cards. Distinguish at least one system boundary. Identify one missing local input. Use calibrated language such as “modeled,” “suggests,” “potential risk,” or “cannot yet determine.” If you consult an AI-generated explanation, pause before accepting its wording. Identify one place it could be wrong, one fact you would check, and one source that could support that check. Your final judgment should be grounded in the evidence cards and the local question. Sources - T. Xiao et al. (2025), Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA, Nature Sustainability, doi:10.1038/s41893-025-01681-y. - N. Lei et al. (2025), The water use of data center workloads: A review and assessment of key determinants, Resources, Conservation & Recycling 219, 108310, doi:10.1016/j.resconrec.2025.108310. - H. I. Jager & H. S. Yoon (2026), Data centers: an emerging threat to freshwater biodiversity in the United States, Water Biology and Security, 100585, doi:10.1016/j.watbs.2026.100585.