Login

For user
For provider
Search Easy Read

AI and sustainability – Can I use AI sustainably?

This image was created using AI. Source: midjourney

 

Artificial intelligence (AI) has made enormous strides in recent years, particularly in the field of so-called generative AI. This includes large language models such as ChatGPT, as well as applications that can generate images, videos or audio recordings, such as Midjourney, Dall-E or Adobe Firefly.

For you as a user, AI models are often handy: for example, you ask a question (formulate what is known as a ‘prompt’) and receive the answer in a fraction of a second – what an incredible feat! What isn’t immediately apparent, however, is that running AI consumes a great deal of natural resources.

This raises a number of questions: How sustainable is AI’s ‘operation’? How much electricity and cooling water are required for AI models such as ChatGPT? Can I use (generative) AI in an energy-efficient way?

What is consumed, and how much?

1. Electricity consumption

Every key you press on your computer or mobile phone keyboard, and every word that appears on your monitor or display, triggers a computational process that consumes electricity. When you use a GPT model – that is, a pre-trained model such as ChatGPT – and ask a question, for example, what you have typed in and the AI’s response are both transmitted via the computer’s electrical signals to a server. However, the providers of large AI models cannot manage this vast ‘information superhighway’ with a single server; instead, they are building entire data centres – that is, large computer and server facilities. OpenAI, the operator of ChatGPT, is planning, for example, a new data centre in the US, complete with a gas-fired power station that is set to be large enough to provide up to 8 gigawatt-hours (GWh) of computing power to run its own AI. This amount of energy could supply 20 million households with electricity for a whole year – equivalent to half of Germany. [1]

When it comes to AI’s electricity consumption, two phases can be distinguished: the training process (training the AI model using vast amounts of data) and the ongoing operation of the AI (its actual use).

Phase 1 (Training): When training language models, an enormous amount of information is fed into the AI system, which later forms the basis of your search results. Example: The training of GPT-3 alone (the version released in 2023) is estimated to have consumed 1.287 million kilowatt-hours (kWh) of electricity. This amount of energy could supply around 400 households in Germany with electricity for a whole year. With newer models, the training methods are becoming even more complex and resource-intensive. However, training is a one-off process: once completed, the finished model can be used repeatedly. [2]

Phase 2 (Use): Ongoing operation requires far more energy than the one-off training. Every time a user submits a query and the model generates a response, thousands of computing units are working in the background. By way of comparison: according to the company, a standard Google search (without AI support) requires only around 0.3 watts per hour (Wh). This amount of electricity would be enough to power a small LED lamp for approximately three minutes. The same query (prompt) on ChatGPT consumes almost ten times as much (2.9 Wh). [3]

2. Water consumption

You’ve no doubt noticed that your laptop or mobile phone gets quite warm when they’ve been running for a long time. Servers in data centres generate heat in exactly the same way, but on a much larger scale. The water required for cooling comes either from cooling towers (where the water is reused) or from cooling systems that use outside air. This can have an impact on the ecosystem in the region where the data centre is located. According to estimates, training GPT-3 alone consumed around 5.4 million litres of water [4]; however, it is difficult to obtain reliable figures, as, on the one hand, companies are reluctant to disclose this information and, on the other, it is not always possible to clearly distinguish which natural resources were required for which process. For example, the production of microchips – which are used in all computers – also consumes a great deal of water. And even more concerning: after production, radioactive waste can remain in the water. [5]

3. CO₂ emissions

Depending on where the electricity for data centres comes from, different levels of emissions may result. When renewable energy sources (e.g. solar or wind power) are used, electricity generation produces significantly lower emissions than when fossil fuels (e.g. lignite or natural gas) are used. Due to the enormous energy consumption of AI models, the continued expansion and use of AI also means that gas- and coal-fired power stations remain online for longer or, as in the case of OpenAI, are even being built from scratch. To reduce CO₂ emissions nonetheless, the expansion of renewable energy would need to be driven forward more vigorously. [6]

This image was created using AI. Source: midjourney

How can I use AI more sustainably in my everyday life?

Use AI in a targeted manner and avoid unnecessary enquiries.

Think carefully beforehand about whether you want to ask your question to a generative AI such as ChatGPT, or whether a standard search using a search engine (e.g. Google, Ecosia or Bing) might be just as suitable. Using AI isn’t always the best solution, even in terms of the results you’re shown (take a look at the explanatory video ‘Why we shouldn’t always believe everything ChatGPT and co. tell us’). You can ask yourself the same question when editing images, for example. Does it have to be AI here, or might a traditional image-editing programme perhaps suffice? (Note: some image-editing programmes also use AI within their software, e.g. Canva).

Reuse results rather than generating them from scratch.

If your query is on a similar topic to a previous one, please use the existing search query and build on it. That way, the AI can reuse the previous answers and results, thereby saving energy.

Choose the right AI.

Not every AI ‘can’ do everything equally well. Carefully selecting the right AI saves resources, as content can be created more quickly when the AI is tailored to that specific use case. There are suitable AI systems for generating images, sounds, videos or text.

This image was created using AI. Source: midjourney

It’s all down to the numbers!

Other digital activities naturally consume electricity too. Comparison with the electricity consumption of other applications (estimates [7]):

  • One AI query = 200 text messages on WhatsApp
  • 7–8 queries to an AI = one hour of music streaming
  • 24 AI queries = one hour of video streaming

It is not the individual private prompt that poses a problem, but the sheer volume of millions of search queries, image generation requests and the like! As many AI models are available free of charge for personal use, are often easy to use, are fun to interact with and potentially offer more information and presentation options than traditional search engine queries, it is easy to lose sight of their environmental footprint.

And, of course, private use of AI is in no way comparable to industrial use of AI, which consumes far more computing power. On the other hand, AI can also help to optimise industrial processes, thereby conserving natural resources. It remains to be seen whether, many years from now, the overall balance will be positive, as (generative) AI can be a key building block for a more sustainable future – provided its use is managed responsibly. Sustainability does not arise from technology alone, but from the decisions of the people who develop and use it.

Date: 15.10.2026
Authors: Linda Zeuner, Jessica Wawrzyniak