Environmental impacts of artificial intelligence

In this article, Matthew Belsey, Environmental Reporting and Data Manager at Culture for Climate Scotland, gives an overview of what artificial intelligence (AI) is and the environmental impacts of this technology.

What is AI?

According to IBM’s ‘What is artificial intelligence?’ guide, AI is ‘technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy’. AI is a catch-all term that covers a multitude of technologies that have been developed and discussed by scientists and mathematicians since the 1950s.

Prior to the AI boom in 2020, when people referenced AI they would generally mean mathematical models (algorithms) that could make predictions or reach outcomes related to a dataset that they’d previously been trained on, without explicit programming. Since 2020, a particular subset of AI, called generative AI, has become increasingly popular and widespread. Generative AI generates text and media based on prompts inputted by a user – eg OpenAI’s ChatGPT, Google’s Gemini and Microsoft’s Copilot. I’ll be focusing on generative AI for much of this article.

It’s important to note that AI is a fast-moving industry with technologies developing all the time. The following information is true at time of writing and we did not use AI to help write this article! We do not use AI to write any of our content, as stated on our sustainability page.

Environmental impact of AI

Culture for Climate Scotland is focused on connecting culture and climate change, therefore it makes sense to first consider the environmental impact of AI. Generative AI is extremely resource-intensive, consuming a lot of water and raw materials, and relying heavily on fossil fuels for its huge energy consumption. We will expand on each of these areas in this section.

Energy

The large language models (LLMs) powering generative AI user interfaces are trained on vast amounts of data to give you an answer or output to the prompt you’ve submitted. One AI text query uses around 0.00003 kWh of energy according to a University of Michigan study highlighted in this MIT Technology Review article. Therefore, 14,000 AI text queries is about the same energy as a day’s laptop use (0.4-0.5KWh). That might not sound like a lot, but ChatGPT receives over 2.5 billion queries per day, which equates to over 75,000 kWh of electricity per day. Over a year this would equate to more energy than 10,000 UK households.

It is worth noting that this energy usage value is an estimate, since this information isn’t publicly provided by big tech companies who own these large AI models and infrastructure, such as Open AI, Google and Nvidia.

The energy ethics of AI are also important to consider. Most of the energy used by AI is not the energy being used by your laptop when you send a query on a generative AI platform, it’s being generated in a data centre located somewhere else. Defined in IBM’s ‘What is a data centre?’ article, a data centre is a physical space containing IT infrastructure (essentially many powerful computers) for running online applications and services. These data centres are located across the world: for example, the Microsoft Datacentres web page shows the location of Microsoft’s data centres.

Some data centres use as much energy as a small city. This means there is the potential for energy shortages in the areas they are located as well as increased use of fossil fuel power plants to power them and increased energy costs in their local areas as highlighted in the Forbes article ‘As AI Booms, Data Centres May Create Electricity Scarcity Among Users’. Data centres are frequently located in highly deprived areas, which poses greater risks to local communities facing these energy shortages and higher energy costs. It is therefore not much of a leap to suggest that the AI industry has grown to reflect a similar global structural set up to the manufacturing industry, where big tech companies are headquartered in high-income countries/areas, but locate their technological infrastructure in low- and middle-income countries/areas. In their Energy and AI report, the International Energy Agency (IEA) highlight that fossil fuels provide nearly 60% of power to data centres. Therefore, even if you are getting renewable energy supplied to your house or office, it’s likely that the energy supplied to the data centre – where your information is held and processed – won’t be.

Water

Along with energy, generative AI also uses a lot of water. Water is primarily used in the data centres to keep the IT infrastructure cool. The water that is used in this process must then be treated before it can be used as drinking water or even for server cooling again.

AI is predicted to be the main contributor to an increase in global water use from 1.1 billion to 6.6 billion cubic metres by 2027 as highlighted in an article on the UK Government Sustainable ICT blog. A single query sent on a generative AI platform like ChatGPT is estimated to use anything from 0.3 ml to 500 ml of water. However, as previously mentioned, this information is also not being regularly disclosed by AI companies which makes it hard to agree on a concrete figure.

As with energy, the water comes from wherever the data centre is. Increasingly, these data centres are being built in areas with existing water stress. An investigation by Bloomberg looking at the impacts of AI data centres on water security found that two thirds of new data centres since 2022 were built in areas already experiencing high levels of water stress,  often in water scarce areas. In their 2023 environmental sustainability report, Microsoft said that 41% of its water used for AI purposes came from ‘areas with water stress’. With a report from the United Nations University Institute for Water, Environment and Health stating that we are now in an era of ‘global water bankruptcy’, protection of our sources of drinking water is more important than ever.

Further reading on this topic:

• How AI uses our drinking water – BBC World Service
• BBC article on the water footprints of Scottish data centres

Raw materials

The third important environmental impact of AI is the use of raw materials. AI data centres are essentially many big computers. Making one 2kg computer requires 800 kg of raw materials according to the 2024 Digital Economy Report from the United Nations Conference on Trade and Development. This includes rare earth elements, which are often mined in environmentally and socially destructive ways as highlighted in the World Resources Institute’s article on ‘The Critical Minerals Conundrum’.

Is AI really that bad?

Let’s explore the above approximate numbers of AI’s energy, water and raw material use in the context of other energy consumptive industries.

The IEA’s 2024 Electricity Report showed that AI data centres were responsible for just under 2% of global electricity in 2024. In their World Energy Outlook, the IEA said that ‘at a global level, data centres account for a relatively small share of overall electricity demand growth to 2030’.

A Forbes article compiling data from a TRG Datacentres report – New Data: AI Is Almost Green Compared To Netflix, Zoom, YouTube – compared AI use with other common online activities:

• YouTube or Netflix, 1 hour (HD) = ~0.12 kWh
• Zoom, 1 hour = ~0.0486 kWh
• AI Text-to-video generation, 6–10 seconds = ~0.05 kWh
• AI image generation, 1 image = ~0.003 kWh
• AI chatbot prompt = ~0.0003 kWh (note the 10-times higher energy amount than the value we discussed earlier in this article highlighting the range of estimations)

While these comparisons give useful context to the demands of AI systems when measured against other electricity-intensive activities, the comparison of streaming Netflix vs using ChatGPT is perhaps not very helpful. They are used for different purposes so you can’t say that doing one thing rather than the other would be better or worse. We’re likely to be using both streaming and AI for different purposes so we should be exploring how we can reduce our impact in both areas.

What we can say for definite is that AI does have an environmental impact. Therefore, it’s imperative that if you are exploring using AI within your organisation, you also consider the wider environmental impact of your use case(s). You could even ask whether using AI is the correct course of action – could you find another solution? Popular discourse encourages organisations to adopt AI or risk being left behind as technology alters our ways of working. It is being framed as a tool to allow us to be more productive and boost operational efficiency, among other things. However, we must ask ourselves whether the gains made by using AI outweigh the releasing of additional CO2e into the atmosphere. Is each AI generated line of text worth an extra 0.1g CO2e, if we could easily spend a bit more time to create it ourselves with a lesser knock-on environmental impact?

AI in wider context

While this article has focused on exploring the environmental impacts of AI, it is worth acknowledging wider concerns around the use of AI. These include the threat to intellectual property, the presence of bias in the data used to train AI models, worries over privacy and security, and not understanding how an AI model makes its decisions (commonly referred to as the ‘black box’ problem). On a global scale there are also concerns around mis/disinformation as explained in CAAD report on The AI Threats to Climate Change, the centralisation of power and the political influence that large AI companies are having.

Equally, there are many wider benefits that AI can offer to the world including advancements in medicine, helping with solving the climate crisis and use in research more generally. However, in each instance, AI must be used ethically, responsibly, fairly and justly.

Here are some resources with further details on these topics:

• Arts & Culture Podcast – AI in Culture
• Responsible AI in practice: The Journey of Arts Council England
• The Guardian – Five ways AI could improve the world
• Nature – The Future of AI
• UNESCO – Ethics of Artificial Intelligence
• State of AI Ethics Report by Montreal AI Ethics Institute (MAIEI)

This article is not intended to scare you into not using AI. I hope it forms part of your reading as you or your organisation make an informed decision about AI use. If you would like to discuss this further or have any questions, please get in touch at [email protected].

Header image: Graphic adapted from a photograph by Sasirin Pamai on Canva.