Environmental Impacts of AI – A very brief primer
By Alasdair Bachell, ARA Environmental Sustainability Group Research Officer
It’s difficult to get away from the world of AI both in working and personal life - you may well already be using it to do any number of tasks from writing emails to generating images. The many uses and grand promises made by its evangelists as to how it will transform the world of work are matched only by its potential for malicious or unethical applications. Questions as to the ethics of AI usage are numerous, especially among information professionals, but we will be concentrating on one in this post: how environmentally sustainable is it?
As with other areas of digital sustainability, this is not a simple question to answer. This short blog will go over some of the basic issues but any one of them can be explored in much greater depth.
One issue that confuses matters is what exactly we are talking about when we say “AI”. Generally, people are talking about something like ChatGPT or Microsoft Copilot, tools that can answer questions and create text or images based on user input. However, the term also refers to a whole host of services and tools, all doing different things and working in different ways. A key differentiator is between analytical AI and generative AI.
Analytical AI describes tools designed to parse large volumes of data, describe these data sets, create and optimise models and use these to predict future trends. These kinds of tools are seeing use in various fields including finance and healthcare and the data it utilises is typically more specialised and curated. Generative AI (GenAI) are tools trained on large data sets to generate new content such as text, images and audio. To do this GenAI is trained on vast quantities of data scraped from all corners of the internet (this is where copyright discussions around AI stem from – a topic for someone else!)
Regardless of the application, AI requires a huge amount of infrastructure, namely data centres. These provide the computational power required to train AI systems. It’s here we can start talking about the impact of AI on energy use. An individual query to an AI model uses a very small amount of energy, with simple questions using very little while image and video generation use significantly more. All together these individual queries can add up to a lot of electricity – around 80-90% of AI’s total energy consumption, with the remaining percentage used for training models.
Before we proceed, it is important to note that many of the available figures on AI resource usage are estimates on the part of journalists and industry analysts. The companies providing AI services keep their actual figures closely guarded. What is certain is that the boom in data centre construction is leading to massively increased energy and water consumption (which we’ll come to shortly).
Data centres accounted for around 1.5% of world’s total electricity consumption in 2024 according to the International Energy Agency (IEA), a figure which is set to double by 2030. This is a relatively modest amount in the grand scheme of things, but a more granular look at where demand stems from shows a more regionally concentrated increase in demand. For example, in the USA data centres are expected to rise to 7-12% of its total consumption by 2028.
Data centres require a stable, and constant, supply of electricity so intermittent renewable systems such as solar and wind can only ever be part of their energy solution. That leaves fossil fuels, nuclear energy, or extensive battery storage to provide that constant supply. This means that the growth in demand will likely still be met by a large proportion of fossil fuel generation. In the USA, a huge spike in new gas generation is being driven by AI data centres, which will vastly increase carbon dioxide emissions.
Apart from energy usage, data centres also require cooling which means that many of them need water – lots of water. Global annual demand for water for AI usage is estimated to be 4.2 – 6.6 billion cubic metres of water withdrawal, higher than the total water withdrawal of Denmark. While water isn’t actively consumed in the process of cooling, it is water taken out of local systems which means less for everyone else. Where this use is from potable water sources, this means less available drinking water and less water for agriculture. In areas already susceptible to drought, this can be a significant pressure on both people and the local environment.
Much of this is uncertain, not least because of the uncertainty around the figures we do have. We do not know how much the appetite for AI data centres will grow. If the current AI bubble bursts will this demand be curtailed or will data centres be shut down entirely? Will changes in public attitudes lead to a decrease in demand for AI tools? More advanced models which require less energy use could be developed. There is speculation that AI could be used to drive down emissions in other sectors, offsetting its own carbon footprint –though the IEA notes“that there is currently no momentum that could ensure the widespread adoption of these AI applications.”
Environmental impacts are social impacts. The growth in energy demand means that the increase in capacity is not being used for other key industries trying to decarbonise. Renewable targets may be under threat from the increased proportion of fossil fuel generators on the grid. Water scarcity is already a huge issue for many parts of the world and new data centres add significantly to existing pressures.
Even if every environmental question was put to rest, the other numerous ethical and legal issues will remain. It is up to us as information professionals to decide to what extent we engage with these systems, to be cognisant of the potential for misinformation or outright harm. For the foreseeable future, we must recognise that, however small, using AI systems does have a measurable negative impact on the environment. We must now decide if it is worth the cost.