Is AI Bad for the Environment? The Honest Answer

Short answer: one AI answer is not bad for the environment, but AI as a whole can be, because of how fast it is growing and how it is powered. A typical text question uses about as much electricity as a few seconds of TV. The problem is that billions of those questions, plus images, video and long documents, are driving a rapid rise in data-center electricity use, much of it still made from fossil fuels.

Whether AI ends up being bad for the planet depends less on any single prompt and more on two choices: what powers the data centers and how efficient the models are.

What one AI answer really costs

The best figures come from the companies that run the models:

	Electricity per text answer	Water per text answer
Google Gemini (median prompt)	0.24 Wh	0.26 mL
OpenAI ChatGPT (average query)	About 0.34 Wh	About 0.32 mL

For scale, 0.24 Wh is what Google compares to watching TV for less than nine seconds, and 0.26 mL is about five drops of water. Independent researchers at Epoch AI put a typical GPT-4o query at about 0.3 Wh, in the same range.

Some tasks cost much more:

• Long documents. The model reads everything you send, every time. Epoch AI estimates a query with a 10,000-token input, a long report, at about 2.5 Wh.
• Images. In a peer-reviewed study of open image models, generating one image used about 2.9 Wh on average, as much as roughly ten text answers.
• Reasoning modes and video. Models that “think” step by step, and video generation, use far more again.

Older figures you may still see quoted, such as 3 Wh per ChatGPT question, are roughly ten times too high for today’s models, according to Epoch AI.

Why the total is the real problem

Per answer, AI keeps getting more efficient. Google reports that the energy of its median Gemini text prompt fell 33 times in the twelve months to May 2025.

But use is growing faster than efficiency. The International Energy Agency estimates that data centers worldwide used about 485 TWh of electricity in 2025, up 17% in a year, with AI-focused data centers growing 50%. It expects total data-center use to roughly double to 950 TWh by 2030, around 3% of global electricity demand.

That growth has to be powered by something. Where it is met with new gas plants or by keeping coal plants open, AI adds real emissions. Several big tech companies have seen their emissions rise as they build out AI; Google, for one, has stepped back from its 2030 net-zero pledge.

Water

Many data centers cool their chips with evaporative cooling towers, and the water they evaporate is lost to the local supply. Power plants that run on coal, gas and nuclear use water too.

Per answer, the amounts are small. In total they add up: researchers at UC Riverside and UT Arlington project that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027. It matters most where data centers are built in hot, dry, water-stressed regions. Read more in which AI uses the least water.

Hardware and training

Two parts of AI’s footprint rarely show up in per-answer figures:

• Training. Building a large model takes a great deal of computing before anyone uses it.
• Hardware. Making chips and servers uses energy, water and minerals, and the hardware is replaced every few years.

Mistral is one of the few companies to publish a full life-cycle study covering both. It found that impacts rise roughly in line with model size, which is one more reason to use smaller models when they are good enough.

What decides whether AI is part of the problem

1. Where the electricity comes from. The same answer has a tiny carbon footprint on a hydro or wind-powered grid and a much larger one on coal. This is the biggest single factor.
2. How efficient the models are. Right-sized models, used for the right jobs, cut energy by large factors.
3. How the data centers are cooled. Cold climates and closed-loop cooling use far less water than evaporative towers in hot regions.
4. Whether new clean energy gets built. AI’s growth can either lock in fossil power or pay for new wind, solar and storage.

What you can do

• Use a smaller, faster model for everyday questions.
• Skip reasoning modes, image generation and video when you don’t need them.
• Attach only the part of a document you need.
• Choose an AI app that says where its data centers are and what powers them. See what makes AI eco-friendly.

How Viro handles this. Viro’s own model runs in data centers in Iceland, Norway, Finland and Ireland, powered by hydro, geothermal and wind, with cooling that relies on cold outside air and closed loops rather than evaporation. When you use ChatGPT, Claude or Gemini inside Viro, those answers run on their makers’ servers; Viro estimates the electricity each message uses and funds new clean energy to match it. Every reply shows an estimate of the electricity it used. See the data centers.

AI does not have to be bad for the environment. Run on clean power, with efficient models, its growth can help pay for the clean energy build-out instead of slowing it down.

Try Viro AI free: ChatGPT, Claude, Gemini and Viro’s own renewable-powered model in one app.

Frequently asked questions

Is AI bad for the environment?

It can be, mostly because of scale and how it is powered. One text answer uses about 0.24 to 0.34 watt-hours, a few seconds of TV, but data centers used about 485 TWh of electricity in 2025 and the IEA expects that to roughly double by 2030. AI run on clean power with right-sized models has a much smaller footprint.

How bad is ChatGPT for the environment?

OpenAI says an average ChatGPT query uses about 0.34 watt-hours of electricity and about 0.32 mL of water. Per question that is small. The impact comes from billions of questions a day and from heavier tasks like images and long documents, and OpenAI has not published what powers its data centers.

Does AI use more energy than Google search?

For simple questions the gap has become small, because AI answers have become much more efficient. The IEA estimates that if every conventional web search were done as a simple AI text query, it would use under 4 TWh of electricity a year, less than 1% of what data centers use today.

Is AI getting better or worse for the environment?

Both. Energy per answer is falling quickly: Google cut the energy of its median Gemini text prompt 33 times in a year. But total use is rising because people use AI much more, so where data centers get their electricity matters more every year.

How can I use AI more sustainably?

Use smaller models for simple questions, skip reasoning modes and image generation when you don’t need them, attach only the part of a document you need, and choose an AI provider that says where its data centers are and what powers them.

Sources

1. Key Questions on Energy and AI, International Energy Agency, 2026
2. How much energy does Google’s AI use? We did the math, Google Cloud, August 2025
3. The Gentle Singularity, Sam Altman, OpenAI, June 2025
4. How much energy does ChatGPT use?, Epoch AI, February 2025
5. Power Hungry Processing: Watts Driving the Cost of AI Deployment?, Luccioni, Jernite and Strubell, ACM FAccT, 2024
6. Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models, Li, Yang, Islam and Ren, Communications of the ACM, 2025
7. Our contribution to a global environmental standard for AI, Mistral AI, July 2025

Cover photo by Taylor Vick on Unsplash