How to Reduce the Environmental Impact of AI Use

Artificial intelligence is transforming everyday life. Generative AI tools now perform tasks ranging from creating social media images and offering relationship advice to analyzing medical scans, writing software and helping banks detect financial fraud. Behind these services, however, are data centers running energy-intensive computations around the clock.

As AI adoption has accelerated, so has its environmental footprint. Data centers currently use about 414 terawatt-hours of electricity each year—roughly 1.5% of global electricity consumption—according to the International Energy Agency. Electricity use increased by an average of 12% annually for five years before rising by 17% in 2025. By 2030, data-center electricity demand is expected to more than double.

Fossil fuels are supplying much of this additional power, while data centers also require significant amounts of water for cooling. That has raised concerns about the impact of AI infrastructure on local water supplies, particularly in drought-prone regions.

For most individuals, generating an occasional AI image or text response represents only a small share of their personal environmental footprint. Still, experts say technology companies must take the lead in reducing AI’s energy and water consumption through more efficient algorithms, improved hardware and cleaner power sources.

Consumers can also make relatively simple choices to reduce the environmental cost of their AI activity. These include using AI only when it is appropriate, selecting smaller models and writing prompts that limit unnecessary computation.

“Individual choices are not meaningless, and some are more powerful than people realize,” says computer scientist Ivana Drobnjak of University College London.

Why AI chatbots use so much energy

Estimating the energy required for a single AI chatbot request is difficult because consumption varies by model, data center, hardware and prompt length. Google estimates that its Gemini chatbot uses approximately 0.24 watt-hours to answer a median-length text query—about the electricity required to watch television for less than nine seconds. The same request uses roughly 0.26 milliliters of water and produces about 0.03 grams of carbon dioxide.

Those figures are small for one request, but they accumulate rapidly when billions of prompts are processed each day. AI models rely on graphics processing units, or GPUs, which generally use more power than the central processing units that handle simpler activities such as email and conventional web searches.

The large language models behind many AI assistants are also computationally demanding. These systems use transformer architecture to study enormous volumes of text and calculate billions or even trillions of parameters. They then generate responses by predicting which words are most likely to come next.

A transformer-based model must repeatedly process the user’s question and the text it has already produced every time it generates another word. That process involves billions of calculations, making longer answers more energy-intensive than short responses.

AI companies say their systems are becoming more efficient. Google reported in 2025 that Gemini’s estimated energy use of 0.24 watt-hours per median-length text prompt was 33 times lower than the previous year’s figure.

Even with those efficiency gains, the sheer volume of AI use has a major cumulative effect. Based on 2025 data from OpenAI, Drobnjak estimated that approximately 3.2 billion requests were being sent to ChatGPT each day. Users are also asking AI systems to create huge quantities of text, images and video, while automated “AI agents” increasingly generate queries of their own.

So how can people reduce the energy and water associated with everyday AI use? Researchers recommend several practical strategies.

Use traditional search when AI is unnecessary

The easiest way to reduce AI energy consumption is to consider whether an AI tool is needed at all. Asking a chatbot questions such as “What should I wear?” or “What is the weather?” requires far more computing power than using a conventional search engine or weather app.

The same principle applies to search engines that automatically generate AI summaries, including Google AI Overviews and Microsoft’s Copilot search. If you only need to find a specific webpage or article, choosing a standard web-results option can reduce unnecessary processing. Some browsers also allow users to disable AI features or add “-ai” to a search query.

Choose smaller AI models for simple tasks

People and businesses that regularly use AI for focused tasks—such as translation, summarization or basic question answering—can reduce energy use by selecting smaller, specialized language models. These models perform fewer calculations than large, general-purpose systems and are often better suited to routine applications.

In a 2025 UNESCO study, Drobnjak compared specialized models, including opus-mt-en-es for English-to-Spanish translation, with Meta’s much larger Llama 3.1 model. Although smaller models may be less convenient than popular commercial chatbots, many are freely available through platforms such as Hugging Face.

The study found that smaller models used between 15 and 50 times less energy while producing higher-quality results on the specific tasks they were designed to perform. Moving from large general-purpose models to smaller alternatives reduced total energy consumption by about 90% in the study.

As AI expert Udit Gupta of Cornell Tech explains, a trillion-parameter model is unnecessary for a task as simple as editing an email.

Ask AI chatbots for shorter answers

Because large language models perform calculations for every word they generate, shorter responses generally require less energy. AI systems researcher Mosharaf Chowdhury of the University of Michigan has found that models that produce especially long, detailed answers tend to consume more electricity.

For example, one version of Alibaba Cloud’s Qwen model uses considerably more energy in its reasoning mode, which produced approximately 10 times more words than its standard text-conversation mode. Users can reserve reasoning modes for complex problems and use standard modes for routine questions.

A simple instruction such as “be brief” or “use no more than five bullet points” can also reduce energy use. In the UNESCO study, asking Llama to cut its response length in half reduced energy consumption by about 50%. By comparison, shortening the prompt itself produced a smaller reduction of about 5%.

“The size of the output is what determines and drives the energy expenditure the most,” Drobnjak says. Her energy-saving recommendations for AI users include writing precise prompts and setting clear limits such as “five bullets maximum.”

Generate images and videos at the lowest useful resolution

AI image and video generation can require far more energy than text generation because the systems repeatedly process millions of pixels. Each new version may involve analyzing and modifying the entire image or video frame.

To reduce the environmental impact of AI-generated media, create images and videos only when necessary and use the lowest resolution that meets your needs. Starting with a low-resolution draft and increasing the quality later can prevent wasted computation. Editing an existing image is also generally less energy-intensive than generating a completely new image.

When creating several images, generating them in one batch or session can be more efficient than submitting many separate requests.

These personal actions will not eliminate AI’s environmental impact, but their combined effect can be meaningful. While technology companies, researchers and policymakers work on more efficient AI systems and cleaner data-center infrastructure, users can help by choosing the right tool for each task.

Editor’s note: This story was updated on July 21, 2026, to clarify that the energy use of individuals who use AI tools frequently is cumulative, not necessarily large, as was originally stated.

This article originally appeared in Knowable Magazine, a nonprofit publication dedicated to making scientific knowledge accessible to all. Sign up for Knowable Magazine’s newsletter.