A colleague told me she wanted to stop using AI tools altogether. She had read about the environmental toll of data centers on low-income communities, and it kept her up at night. She did not want any part of it, despite knowing how these tools could help her in clinic.
I understood where she was coming from. I also did not think the answer was to look away. So I asked two of my dear friends, colleagues and experts in this area, to sit down and talk it through with me.
Dr. Manijeh Berenji is an occupational and environmental medicine physician who also works in clinical informatics. She spends her days at the seam between what people are exposed to and how it shows up in their bodies. Betty Villantay is a first-year medical student at Nova Southeastern University with an engineering background and a long-standing interest in environmental health. One of us is deep in practice. One of us is entering it. All three of us use these tools, and all three of us have questions about what they cost ethically, environmentally, and economically.
This post is the companion to that conversation. The recap is above in the recording. Below is what stayed with me, plus the data and one practical thing you can change today.
What the conversation kept returning to
The main idea was not “AI is good” or “AI is bad.” It was resource consumption. We do not yet have honest numbers for what these tools cost, and you cannot manage what you cannot count.
Dr. Berenji opened with something concrete. In her own county, the city of Monterey Park recently put a data center to a ballot measure, and residents rejected it. That is not an isolated report. Recent Gallup and Pew polling she pointed to suggests a large majority of Americans do not want these facilities near them, mostly over noise, water use, and air quality. Communities are learning what a data center is before the healthcare system has finished deciding how it feels about one.
Betty came at it from a different perspective. She described being honest with herself about prompts. A simple question costs very little. But feeding in a full lecture slide and asking the model to analyze images and generate a diagram costs much more, and it adds up across millions of students doing the same thing at the same time. Her point was not guilt. It was awareness, and then better habits. She continues to hand-draw her study aids and diagrams, and she thinks harder before she presses enter.
Dr. Berenji connected it to the room I care most about: the exam room. Her health system recently rolled out an ambient AI scribe, and clinicians find it genuinely helpful. Then a colleague asked her a question she could not answer. How much energy does one ambient scribe clinical encounter actually use? She did not know. Most of us do not have that answer yet.
This is where I want to point to a colleague and friend whose work deserves more attention. Dr. Chethan Sarabu is a pediatrician and clinical informaticist who directs clinical innovation at Cornell Tech’s Health Tech Hub, and he trained in landscape architecture before medicine, which tells you how he thinks. He and his collaborators recently published a decision framework called Sustainably Advancing Health AI, or SAHAI, in NEJM Catalyst. This is particularly important as it gives health systems a way to measure the energy, emissions, and cost of specific clinical uses and to choose the lighter option when one exists. He used the EHR inbox as a starting unit: how much energy does it take for AI to answer one patient message, multiplied across a day, a clinic, a system? Not “AI is heavy,” but “this task, this many times a day, costs this much, and here is the lower-cost way to do it.” If you take one citation from this post into a meeting at your own organization, make it his.
The part that will stay with me came from patients. Dr. Berenji and I both described younger patients walking into the clinic and asking, unprompted, about the water and energy behind the AI tools we use. The people we serve are already thinking about this. The least we can do is be able to answer them.
I also loved what Betty shared about the cycle of life. She trained as an engineer before medicine, and kept returning to a simple principle from that world: what you build should be able to return back to the earth. She talked about “green chips,” the same output for less energy, and about designing data centers that work with soil and water and trees instead of against them. Her recommendation is to slow the pace of building long enough to ask who pays for it.
What the evidence actually shows
Here is the data underneath the conversation, organized so you can use it. Full citations are at the bottom.
Energy. Data centers used about 1.5% of the world’s electricity in 2024, roughly 415 terawatt-hours. The International Energy Agency projects that doubles to around 945 TWh by 2030, driven largely by AI. In the United States, data centers could reach 6.7 to 12% of national electricity demand by 2028.
Carbon. A 2025 Nature Sustainability analysis estimates the AI industry could add 24 to 44 million metric tons of CO2 per year by 2030 on its current path, comparable to putting 5 to 10 million more cars on US roads. The demand is real enough that it is delaying coal plant retirements and pulling new natural gas online.
Water. Cooling is thirsty. Depending on design and climate, data centers evaporate roughly 1 to 9 liters of water per kilowatt-hour of server energy, and a mid-sized facility can use on the order of 110 million gallons a year. In some cases more than half of that cooling water comes from municipal drinking supplies. Since 2022, nearly two-thirds of new US data centers have been built in places already under water stress.
Hardware. The fast turnover of specialized AI chips is generating a growing e-waste stream, potentially millions of tons this decade, carrying heavy metals like lead and mercury that often end up processed in the Global South.
Who carries it. This is the part that belongs to each of us. One 2024 analysis found the public health burden from data center emissions is deeply unequal, with per-household health costs in the most affected communities reaching far higher than in less-exposed ones. These facilities are disproportionately sited near low-income neighborhoods and communities of color, and near the fossil fuel plants that power them. The exposures are familiar to any primary care clinician: fine particulate matter that worsens asthma and COPD, cardiovascular strain, and chronic noise that disrupts sleep.
However, not all is negative; there are some benefits too. These projects bring tax revenue and well-paying jobs, and the tools they power can genuinely help patients and clinicians. The honest position is not refusal. It insists on conducting the environmental impact assessments, reviewing the community benefit agreements, and determining the short- and long-term risks (including measurements), before the concrete is poured, and advocating for more environmentally friendly technology.
What we still do not know is as important as what we do. Facility-level data on water use, energy sourcing, and backup-generator emissions are largely hidden or unavailable. Long-term studies tracking health outcomes in communities that host these centers are still scarce. This is a developing area of research.
One thing you can change today: search without the AI layer
This came up near the end of the conversation, and several of you will find it useful. A plain web search uses far less energy than generating an AI summary for the same question. If you only need a link or a quick fact, you do not need the model to write you a paragraph. That uses more resources than necessary.
Google no longer offers a single official “off” switch for AI Overviews, but there are reliable ways to get the classic list of links.
The quick way (any device). Run your search. Under the search bar, look at the row of filters (All, Images, News, and so on). Click Web. If you do not see it, click More first. Your results come back as plain links, no AI summary. You repeat this each search.
The permanent way (Chrome or Edge on a computer). Make the plain “Web” view your default so you never have to think about it again:
Open your browser settings and go to Search engine, then Manage search engines and site search.
Under Site search, click Add.
Fill in the fields. Name it something like Google Web. Shortcut: web. In the URL field, paste: https://www.google.com/search?q=%s&udm=14
Save it, then click the three-dot menu next to your new entry and choose Make default.
The udm=14 at the end is what forces the plain, link-only view. From then on, your normal searches skip the AI summary automatically.
The no-setup way. Bookmark this and search from it when you want clean results: https://www.google.com/search?q=YOUR+SEARCH&udm=14 (replace the words after q=, using + for spaces). On mobile, the Web filter is the simplest option, since you cannot change search settings as easily.
None of this is anti-AI. It is matching the tool to the task.
Use the model when you need reasoning.
Use a plain search when you need a link.
If you care about this but feel powerless
Betty’s answer to that was the best one in the conversation: Start at home. Your city council meets in public and votes on things like this, often with little turnout. Read your city or county’s environmental and public health reports; they exist, and they will tell you where your community already stands. Show up to one meeting and ask one question: where is the environmental impact assessment? Local nonprofits and online communities are doing this work and will welcome the help. Advocacy scales from the individual, to the grassroots, to the state and federal level. It rarely starts at the top.
For those of us in healthcare, Dr. Berenji’s motto is a good one: use these tools intentionally, ask your organization how it is sourcing the energy behind them, and push for the measurement frameworks that let us weigh benefit against cost with actual numbers instead of vibes.
I would love to hear how this sits with you, especially if patients have started asking you the water question too. Reply in the comments and let me know.
The people in this conversation
Manijeh Berenji, MD, MPH is an Associate Clinical Professor of Medicine at the University of California, Irvine School of Medicine and Wen School of Public Health. She is double board-certified in Occupational and Environmental Medicine and Preventive Medicine, and board-eligible in Clinical Informatics. She works at the intersection of workplace and environmental exposures and how technology shapes care. She has been educating peers on AI’s environmental footprint through webinars and clinical informatics venues.
Betty Villantay is a first-year medical student at Nova Southeastern University in Florida with a background in engineering and a focus on environmental health and sustainable technology.
Work worth following
Dr. Chethan Sarabu and colleagues published Sustainably Advancing Health AI (SAHAI): A Decision Framework to Mitigate the Energy, Emissions, and Cost of AI in NEJM Catalyst (2025). If your organization is adopting AI tools, this is the framework to bring into the room. Dr. Sarabu directs clinical innovation at Cornell Tech’s Health Tech Hub and co-founded CHILL, the Climate Health Innovation and Learning Lab. Framework: https://catalyst.nejm.org/doi/abs/10.1056/CAT.25.0125 · Plain-language summary: https://tech.cornell.edu/news/health-ai-carbon/
References and further reading
International Energy Agency. Energy and AI (2025). https://www.iea.org/reports/energy-and-ai
Han, Y., et al. The Unpaid Toll: Quantifying and Addressing the Public Health Impact of Data Centers (2024). arXiv. https://arxiv.org/abs/2412.06288
Xiao, T., et al. Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nature Sustainability (2025). https://www.nature.com/articles/s41893-025-01681-y
National Academies of Sciences, Engineering, and Medicine. Implications of Artificial Intelligence–Related Data Center Electricity Use and Emissions: Proceedings of a Workshop (2025). https://doi.org/10.17226/29101
Privette, A. P., Barros, A., & Cai, X. Data Centers Water Footprint: The Need for More Transparency. AGU Advances (2026). https://doi.org/10.1029/2025AV002140
Environmental and Energy Study Institute. Data Centers and Water Consumption (2025). https://www.eesi.org/articles/view/data-centers-and-water-consumption
Mytton, D. Data centre water consumption. npj Clean Water (2021). https://doi.org/10.1038/s41545-021-00101-w
Climate change and health: the next challenge of ethical AI. The Lancet Global Health (2025). https://doi.org/10.1016/S2214-109X(25)00124-X
Wang, P., et al. E-waste challenges of generative artificial intelligence. Nature Computational Science (2024). https://doi.org/10.1038/s43588-024-00712-6
United Nations Environment Programme. AI has an environmental problem. Here’s what the world can do about that (2025). https://www.unep.org/news-and-stories/story/ai-has-environmental-problem-heres-what-world-can-do-about
NAACP. Stop Dirty Data Centers (2026). https://naacp.org/campaigns/stop-dirty-data-centers
Tao, Y., & Gao, P. Global data center expansion and human health: A call for empirical research. Eco-Environment & Health (2025). https://doi.org/10.1016/j.eehl.2025.100157
Sarabu, C., et al. Sustainably Advancing Health AI (SAHAI): A Decision Framework to Mitigate the Energy, Emissions, and Cost of AI. NEJM Catalyst Innovations in Care Delivery (2025). https://catalyst.nejm.org/doi/abs/10.1056/CAT.25.0125 (Plain-language summary: https://tech.cornell.edu/news/health-ai-carbon/)
Disclaimer: All views expressed are my own and do not represent my employer or any institution I am affiliated with. Any tools, products, or technologies mentioned are included for educational purposes only and are not sponsored or endorsed. Nothing in this piece should be interpreted as medical advice.
