How should public interest attorneys think about the environmental costs of AI?

I use AI to help close the access to justice gap. But many public interest lawyers and advocates are skeptical of its potential downsides, so I have found it helpful to communicate with them to understand those concerns. When I wrote “No Devil at this Crossroads,” I chose to focus primarily on the substantive risks and benefits of AI. Since then, I have given many AI trainings and talks around the country, and the most persistent concern that I hear from public interest legal advocates is about the potential environmental impact of using large language models (LLMs), a topic that I only spent a few sentences on in that essay. [A vocabulary note: I will use “AI” to describe the specific technology of LLMs through the rest of this post].

I am not a climate or environmental scientist, but the best information that I have shows that while LLMs are energy intensive and their use is growing, their impact is proportional to their benefits and lower than typical leisure and alternative business activities we engage in without thinking. Datacenters concentrate the negative impacts of AI consumption, sometimes into environmental justice communities, and should be carefully sited.

This post is public interest lawyer-to-lawyer or other advocate in a legal aid program, law clinic, or public defender’s office. I have gathered a wide range of sources and perspectives on the relative impact of AI here, and linked to primary sources that you can explore on your own. In my experience, these are the questions most lawyers want to know the answer to, and nobody else has answered them in one place. I do not claim to be comprehensive. I welcome additional feedback and perspectives to add here.

I tell this story mostly in charts, but maybe the most interesting one is an interactive comparison tool. You can decide what daily activities and level of AI use you want to understand.

Top takeaways

1. Even if you are a heavy user, AI is not a significant part of your personal energy or water use footprint. It is truly miniscule on both counts.
2. AI’s aggregate environmental impact, while measurable, is relatively low.
3. AI datacenters are not a limit on the growth of AI, because AI inference is getting more efficient over time. The story of datacenter growth is primarily a competitive one. Datacenters can concentrate negative impacts and it’s worth thinking carefully about where to site them.

I have a limited ability to make choices to influence my environmental footprint. AI is simply nowhere near my list of concerns when I consider other choices I can make every day, like walking or cycling; or in the long run, like switching to a heat pump and rooftop solar.

Individual AI use is a small part of our daily energy consumption

AI uses a tiny amount of electricity compared to the things we do every day.

Chart: Energy consumption per ChatGPT query compared with everyday electricity use

Chart: Daily electricity scale for AI queries and everyday uses

Chart: Daily electricity scale inset

What about newer reasoning models? Newer AI models often use a trick called “thinking” or “reasoning” that creates a chain of thought that helps the model make fewer mistakes. Estimates are that “thinking” AI models, like GPT-5, can use between 10 and 100x as much energy. We use a conservative estimate of 30x in these charts.

Individual AI is a small part of our daily water use

Some early studies discussed AI as using an unusually high amount of water to do its work. As funny as the memes it spawned are, AI simply does not use an unusual amount of water compared to any other computing task. Most of the water use comes from water used in power generation, so its relative use is described well in the charts above. But some water use comes from cooling datacenters.

Chart: Daily water-use comparison inset

Chart: Daily water-use scale

Aggregate AI use is a growing but still small part of the overall energy sector

Datacenter construction is growing rapidly, although new datacenters are generally more efficient than older ones. This means that datacenter’s share of the global electricity market is growing, too. However, it still is projected to remain in the single digits as a percentage of all energy use through 2030, and currently uses only 1.5% of global electric demand. AI, broken, out, represents only about 32% of datacenter energy consumption.

Chart: Global electricity use — datacenters versus all other uses

Global datacenters used about 1.5% of global electricity in 2025; the IEA projects about 3% by 2030.

International Energy Agency

Chart: U.S. electricity use — datacenters versus all other uses

Chart: AI-focused share of data-center electricity

Training new AI models uses less energy than daily use

Most of AI’s energy impact is from daily users, not training: about 60%. The percentage going to training is projected to continue to slowly decline as AI becomes more common and as training becomes more efficient.

Chart: Estimated inference versus training energy

Daily AI use dwarfs training: at least 60/40 but up to 80/20 in some estimates

Patterson et al. (2022) reported roughly 60/40; Wu et al (2024) report 80/20 inference to training ratio.

Aggregate AI use is a tiny part of global water use

Datacenters, in general, and AI in particular, use a small amount of aggregate water relative to other uses. Datacenter freshwater withdrawals (including both AI and non-AI uses) overall represent about 0.17% of global water withdrawals (calculated against global annual water withdrawals of about 4,000km³). However, localized impact can be greater when datacenters are constructed in water-limited regions. About 72% of AI’s water usage is estimated to come from electricity generation, which means that switching to energy sources like wind and solar can have an outsized impact on reducing AI’s environmental footprint, reducing both carbon and water footprint.

By far the largest consumer of water nationally is agriculture (about 70%), including uses like growing corn for ethanol and almonds in drought-stricken California. U.S. ethanol production alone, for example, uses as much as 17x the water that all global datacenters do. Almonds grown in California exceed global AI datacenter consumption by 10x.

Chart: Selected U.S. water uses

Growing corn for ethanol in the US alone uses up to 17× as much water annually as all global datacenters.

Approximate comparison: 9.3 km³/year versus 0.56 km³/year. These are different geographic and methodological estimates, so “roughly” matters.

AI provider incentives are aligned with better efficiency, and that is happening in practice

AI vendors, especially those who have existing businesses and shareholders, have every incentive to make their models more efficient, because energy use is directly tied to cost. Over just one year, Google was able to cut the energy use of their flagship model by 33x. Read the Google Gemini study. A separate Microsoft Research/Joule analysis shows less dramatic, but still impressive efficiency gains of 10x over time; read the Microsoft study. Hardware is part of the story, too. Holding a model constant but using newer hardware, a Michigan study showed that on average, using the latest computer equipment let vendors use 35% less energy.

Consumers of AI, especially companies with many users, may also use techniques like student-teacher distillation, which trains a small, efficient model to do the work of a large model as it is used over a short period of time, or simply choose small, efficient, general purpose models when they are good enough to do a task.

Chart: Median Gemini text-prompt energy over time

Datacenters have a more complicated story

In theory, putting all AI use into datacenters is a story of efficiency and lowered energy use: big datacenters use less energy per-query than smaller datacenters, and orders of magnitude less than the same calculations made at home, because of economies of scale. However, the tradeoff is that the negative impacts of that construction are concentrated in one place, and that one place can be one that is already overburdened with negative environmental consequences.

Not all datacenters are alike

What’s more, some datacenters are much better stewards of the environment than others. For example, a Berkeley study showed a 10,000-fold range in water consumption for the same tasks across different datacenters. Some datacenters rely on polluting diesel generators while others are more efficient. Because datacenters must run 24 hours a day, it’s common for them to have a backup power generation capacity, which may be relatively dirty.

Datacenter growth is driven by competition, not the need to meet growing AI demand

Princeton Professor Arvind Narayanan argues that datacenter growth represents competitive behavior, but that datacenter moratoria would not meaningfully slow AI’s growth. He suggests that companies compete to build AI datacenters to deprive their competitors of access to compute, a resource that the largest AI companies need to meet expectations and drive profit. On the other hand, AI’s continued efficiency gains would be enough to raise capacity only slightly slower, even if it’s not to the AI vendors’ liking.

Datacenters are sometimes sited in environmental justice communities

An environmental justice community is a neighborhood that is overburdened with negative environmental impacts, such as heavily polluting power plants or, increasingly likely, a datacenter, which may come with its own backup power load. The Urban Institute has created a checklist that they advocate states use to help guide questions for datacenter builders and to ensure that datacenters come with clear community benefits and avoid the worst negative consequences.

Not all datacenters are bad neighbors

Some datacenters are described as having a persistent hum, and when they switch to backup power, they can be especially noisy. But a nonpartisan study in the Commonwealth of Virginia (which houses many datacenters for the company Amazon Web Services) found that only 10% of its datacenters led to noise complaints. Partly this is a question of siting far enough away from residential use, and partly this is a question that can be solved through construction techniques and limits on use of fans and backup power. Datacenters, overall, are relatively quiet neighbors compared to other industrial uses, but they run 24 hours a day.

How do we square concentrated harms against diffuse datacenter benefits?

Datacenters are a story of diffuse benefits (everyone who uses AI) and concentrated harms (concentrated noise, water use, and energy use in the communities where datacenters are sited). This is the kind of harm that can be solved best with uniform regulation that protects residential neighborhoods from noise, places limits on dirty power generation, and requires datacenters to subsidize the cost of added power generation rather than allow utilities to pass on costs to ordinary ratepayers.

Conclusion

I understand that environmental concerns for using AI may be a proxy for broader unease at using AI for daily tasks: concerns that might include labor displacement, concentration of power, surveillance, and other externalities of AI use, ones that could also harm our clients if realized. I discuss some of these questions in more depth in No devil at this crossroads. Not everyone sees value in using AI. For me, I see benefits from using AI in simple, safe ways, almost entirely to help as part of larger tasks. We don’t all have to come out to the same decision. But hopefully it’s helpful to have these figures and references when you are making up your own mind.

Sources and methodological notes

• Epoch AI: How much energy does ChatGPT use? — GPT-4o query estimates and everyday-use comparison spreadsheet.
• Epoch AI comparison spreadsheet — underlying everyday electricity assumptions.
• OpenAI GPT-5 System Card and GPT-5 for developers — GPT-5 fast/reasoning variants and reasoning-effort controls; no public watt-hour-per-query measurement.
• Knoop & Holtmann: Private LLM Inference on Consumer Blackwell GPUs — measured local-GPU inference rates used for the Small, Medium and Large presets; Large uses the available 27B model result.
• Frontier model size estimates — an independent, inspectable estimate puts GPT-5.6 Terra at about 2.1T total parameters. On that scale, the 27B Large benchmark is about 1.3% (roughly 1/78) by parameter count only; this is a guesstimate, not a vendor disclosure or an energy estimate.
• Google DeepMind: Gemini 3.1 Pro model information — official model information for the frontier model used as a comparison point; Google does not publish its parameter count.
• OpenAI: GPT-5.6 Terra model documentation — official model information; OpenAI does not publish its parameter count.
• Google Cloud: Measuring the environmental impact of AI inference — Gemini prompt estimate.
• Mistral AI: contribution to a global environmental standard — Mistral self-reported lifecycle estimate.
• Jisc: Artificial intelligence and the environment — comparison estimates for ordinary online activities.
• IEA: Key Questions on Energy and AI — global data-center electricity estimates and projections.
• Our World in Data: Data centers and AI energy use — AI-focused facility share derived from IEA and Ember data.
• LBNL: United States Data Center Energy Usage Report: 2025 Update — U.S. data-center energy and direct water-consumption estimates.
• USGS: Water Use Across the Conterminous United States — withdrawals and consumptive-use data.
• U.S. GAO: Energy-Water Nexus and U.S. Energy Information Administration — ethanol-production and process-water context.
• Patterson et al.: The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink — older Google ML workload sample from 2019–2021, approximately 60% inference and 40% training; not a universal current split.
• Michigan Engineering: AI energy use — new tools show which model consumes the most power and why; see also the accessible Michigan Engineering News version — measurements of open-weight models and the finding that reasoning models can generate 10–100 times more tokens in some tasks.
• Microsoft Research: Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling — frontier-scale inference estimate and reasoning/test-time-scaling analysis.
• Xiao et al., Nature Sustainability: Environmental impact and net-zero pathways for sustainable AI servers in the USA — U.S. AI-server energy, water-footprint, and siting scenarios.
• ML.ENERGY: Diagnosing Inference Energy Consumption with the ML.ENERGY Leaderboard v3.0 — longitudinal inference-efficiency measurements.
• Lawrence Berkeley National Laboratory: Avoiding Waste Heat through AI Infrastructure Thermal Integration — data-center efficiency and waste-heat context.
• U.S. Department of Energy FEMP: Energy Efficiency in Data Centers — data-center energy-efficiency context.
• U.S. EIA: 2020 Residential Energy Consumption Survey, average site electricity consumption — Northeast household-use benchmark.
• U.S. DOE/Fraunhofer: Urban Luminaire and Display Energy Study — display and monitor power assumptions.
• ENERGY STAR: Television key product criteria and ENERGY STAR: Dishwashers — appliance comparison assumptions.
• U.S. DOE: Conventional Ovens Test Procedure — electric-oven comparison assumption.
• Environmental Challenges (2023): Environmental burdens analysis of current music player devices through LCA methodology — phone/music operational comparison assumption.
• U.S. EPA WaterSense: Showerheads — labeled shower-flow benchmark.
• USDA NASS: Noncitrus Fruits and Nuts 2024 Summary — California almond bearing acreage.
• UC ANR: Almond water use and evapotranspiration — approximately 40–42 inch mature-almond evapotranspiration benchmark.
• USDA Agricultural Research Service: Chapter 11, Water Use and Availability — 2025 corn-ethanol water-use range and irrigation share.
• U.S. EIA: Biofuels and the environment — U.S. biofuel production context.
• Hu et al.: Teacher-Student Architecture for Knowledge Distillation — A Survey — source linked in the article for model-efficiency context.
• UNESCO source document — linked background source on AI and environmental impacts.
• Urban Institute: 25 questions states and cities should ask during a moratorium and data-center break — community-impact and siting checklist.
• Virginia JLARC: Data centers in Virginia — noise-complaint context.
• Scientific American: Colossus data-center community backlash — community-impact reporting.
• LinkedIn discussion on data-center moratoria — commentary cited for a countervailing view on AI growth and efficiency.
• Epoch AI: Optimally allocating compute between inference and training — 2024 analysis arguing that training and inference can rationally receive comparable compute.
• Epoch AI: Most of OpenAI’s 2024 compute went to experiments — estimates of training, inference, and research compute spending.
• Wu et al., Meta: Beyond Efficiency: Scaling AI Sustainably — lifecycle energy and carbon analysis across AI training and inference.