AI's energy costs, water consumption and carbon emissions are rising in step with the growing popularity of chatbots and other AI tools. Finding ways to reduce the energy demands will be important for containing the environmental impacts of the technology.
(Image credit: Witthaya Prasongsin via Getty Images)
Inside the world’s newest data centers, energy-guzzling computations proceed round the clock as AI chatbots and other generative AI tools tackle tasks from the frivolous to the weighty: assembling imagery for social media, proffering relationship advice, analyzing medical images to diagnose cancer, creating code for developers or detecting financial scams for banks.
Just as the popularity of AI tools has skyrocketed in recent years, so have the associated environmental costs. Data centers now consume 414 terawatt-hours per year, or about 1.5 percent of global electricity use, according to the International Energy Agency — an amount that grew by 12 percent annually for five years before jumping to 17 percent in 2025. By 2030, the agency projects that demand for electricity by data centers will more than double. Much of the increasing demand for electricity is being met by fossil fuels, while experts also worry about the use of local water resources to cool data centers in drought-struck regions.
Use of AI to generate text or imagery probably accounts for a mere sliver of any given person’s environmental footprint. And experts stress that the onus is on tech companies to reduce AI’s resource consumption, from creating smarter, energy-saving algorithms to building more efficient hardware.
Yet there are simple actions people can take to ensure that their AI usage has as little environmental impact as possible — from carefully considering where AI is needed to tailoring prompts to minimize the amount of computation required.
“Individual choices are not meaningless, and some are more powerful than people realize,” says computer scientist Ivana Drobnjak of University College London.
It is notoriously tricky to estimate the energy expended on processing an individual chatbot query. Google, for example, estimates that its chatbot Gemini consumes around 0.24 watt-hours to respond to a median-length text query — equivalent to the electricity needed to watch TV for less than nine seconds. It also uses about 0.26 milliliters of water and emits the equivalent of 0.03 grams of carbon dioxide (driving a gas-powered car for a mile would emit about 400 grams). Small individually, these expenditures build up for those individuals and companies that use AI tools a lot.
The reason AI models consume so much energy lies partly in the processors that power them, such as the graphic processing units (GPUs) that consume significantly more energy than the central processing units (CPUs) that fuel simpler tasks like web searches and email. It also has to do with the models that underlie most popular generative AI tools, including the large language models (LLMs) that power AI chatbots and assistants.
Sign up for the Live Science daily newsletter now
Get the world’s most fascinating discoveries delivered straight to your inbox.
These are based on a particular design called transformer architecture. This allows LLMs to train on vast swaths of language patterns in text and, from this, compute hundreds of billions or trillions of parameters. These parameters can then be used to generate new strings of text, by predicting which words are likely to follow one other.
A transformer-based LLM is computationally intensive because for each new word it generates in response to a user’s query, it runs the query and the words that have been written so far through the model, performing billions of calculations each time.
Tech companies note that LLMs have become more energy-efficient over time; according to Google’s 2025 calculations, the 0.24 watt-hours that Gemini consumes on a median-length text prompt represents a 33-fold decrease compared with the model’s energy consumption the previous year.
In any case, even small amounts of energy add up quickly given the scale of AI use. Based on 2025 numbers from tech company OpenAI, Drobnjak estimated in May that, at that point, around 3.2 billion queries are being sent every day to its chatbot ChatGPT. Users are asking AI tools to process and produce vast quantities of text, images and video. Some are having lengthy conversations with chatbots. And, increasingly, people are creating their own “AI agents” that themselves send queries to AI chatbots.
So what can users do to minimize the resources spent on their AI use? Experts have some tips.
Don’t give up on search
As a first, simple measure to save energy, users should carefully consider whether they truly need AI for a given task. “Asking ChatGPT ‘What should I wear today?’ or ‘How is the weather?’ is like taking a Concorde to travel to your supermarket,” says Günter Klambauer, an AI expert at Johannes Kepler University in Austria.
The same goes for web search engines that use AI to automatically generate a response to a query alongside the actual search results, such as Google’s AI overviews or Bing’s Copilot search. “If you’re just looking for a particular article, turning that off could be powerful from a saving-energy perspective,” says Udit Gupta, an expert in electrical and computer engineering at Cornell Tech in New York City. Selecting “Web results only” in one’s browser or including “-ai” in the wording of your web search can do the trick.
Smaller models use less energy
People and businesses that use AI tools a lot for specific tasks like translating or summarizing could consider shifting to smaller language models that are specialized to these tasks. Because these are trained more narrowly and perform fewer computations, they expend less energy than the massive, all-purpose LLMs on the same tasks.
In one 2025 study published by UNESCO, Drobnjak tested the benefits of using small models — such as one called opus-mt-en-es for English-Spanish translations, and other models for summarization and query-answering — in lieu of the model Llama 3.1 developed by Meta. Though these smaller models are often less user-friendly than more popular AI models, they’re freely available from the AI platform Hugging Face. The small models consumed between 15 and 50 times less energy while producing higher-quality outputs on the tasks for which they were designed, the study found.
Using small, specialized models for particular tasks consumes a fraction of the energy guzzled by large, all-purpose models, with similar if even slightly better accuracy.
(Image credit: Knowable Magazine)
The shift away from larger models resulted in a 90 percent decrease in energy use overall, making this the most powerful single energy-saving strategy Drobnjak tested in her study. As Gupta puts it, “You don’t need to use a trillion-parameter model for editing an email.”
Less chatty chatbots
Because LLMs perform so many computations for every consecutive word they produce, it helps to choose models that produce less text in general. AI systems expert Mosharaf Chowdhury of the University of Michigan, who has been measuring the electricity usage of LLMs that have been made publicly available, has learned that models that are “chattier” by nature tend to consume more energy.
For instance, one version of the model Qwen developed by Chinese company Alibaba Cloud consumes significantly more energy when it’s in “problem solving with reasoning mode,” where it produced roughly 10 times as many words in response to a prompt compared to its “text conversation” mode. So some experts recommend using reasoning mode only for complex questions and otherwise sticking with a chatbot’s standard mode.
Simply asking AI chatbots to “be brief” or giving them a word limit can also save energy. In the UNESCO paper, Drobnjak and her colleagues found they could reduce the energy consumption of the Llama model by 50 percent when they instructed it to halve its output. By contrast, keeping the prompt itself short had less significant savings — just 5 percent for a prompt that was half the length of the original query.
“The size of the output is what determines and drives the energy expenditure the most,” Drobnjak says. She has collaborated with the city of San Francisco to develop energy-saving tips for AI users, which include being as specific as possible and adding instructions like “five bullets max.”
Keeping chatbot prompts short can conserve some energy, but asking chatbots to keep their responses brief amounts to much bigger savings.
(Image credit: Knowable Magazine)
Go low-res and batch video
Similar recommendations apply for generating images and video, which can consume orders of magnitude more energy than generating text, as they involve iterating millions of pixels many times over, each time processing the entire image anew, says Drobnjak. Such tools are highly popular: Nearly 40 percent of teens ages 13 to 17 surveyed in a recent study by the Pew Research Center use AI to create or edit images or videos.
Drobnjak recommends generating images or videos only when necessary and only at the resolution necessary. “One option is to just start in low resolution,” she says, “and if the algorithm is in the right direction, you then start increasing resolution.” She also notes that editing existing images is always less computationally intensive than generating new ones from scratch.
And when generating multiple images, it helps to do so in a single session or batch, which is more efficient than doing so in multiple separate requests.
These actions may seem like a drop in a bucket, and in many respects they are, experts say. But small things add up. While waiting for tech companies, scientists and policymakers to find ways of reducing AI’s overall environmental impact, “the individual who knows to reach for the right tool can make a real difference,” Drobnjak says. “AI is [consuming] so much energy that we have to look at it from every angle.”
Editor’s note: This story was updated on July 21, 2026, to clarify that the energy use of individuals who use AI tools a lot is cumulative, not necessarily huge, as was originally stated.
