It’s part of my job as an educational developer at Lund University to talk about Generative AI (GenAI) tools and their place in education. The university provides a couple of GenAI tools for all staff and students, and we need to explain what they do and what impact they may have on teaching, learning, and examination (assessment). Something I hear often from colleagues is that “We shouldn’t use GenAI because it is an environmental disaster.”
I think there are plenty of things to consider when deciding whether or not to use GenAI tools: accuracy, fairness, bias, integrity, security, data privacy, and of course the environmental impact is another very important factor. But I don’t hear that statement in relation to the use of other digital tools. Are GenAI tools a lot worse? I have tried to find out, but it isn’t easy for someone who is not a specialist in sustainability or environmental impacts.
GAI tools (and all other AI tools) are trained by comparing vast quantities of data (non-technical explanation, but I think it is sufficient for this purpose!). This needs a lot of computing power, which in turn means a lot of cooling of computer equipment, which uses a lot of water (Ekin, 2019), as well as increased environmental impact from the manufacture and disposal of the computers themselves. The Organisation for Economic Cooperation and Development (OECD) has recommended a set of approaches to measure and monitor the costs of AI (OECD, 2022), to help governments with policy-making. That could be useful in the longer term to provide a basis for data for non-specialist people like me, but I can’t use that kind of tool myself.
Bashir et al (2024) make a really good case for systematic and targeted development of GenAI so that the selected areas for further work are worthwhile, and not just novel. Unfortunately, research in this area is mostly funded by for-profit organisations, so voices asking for selective development may not be heard very strongly. Commercial organisations are also reticent about sharing relevant data about the energy and water use of their GenAI systems, and this lack of transparency makes it difficult to calculate the environmental impact of both training the models and making individual requests. And whilst we know that the training part uses a lot of computing power, we don’t know much about general use of these tools to generate outputs to regular queries.
During 2025, we have started to get some estimates of impact, based on informed guesses by researchers. Jon Ippolito (2025) has made a set of nine “guesses based on incomplete and often contradictory sources.” which tries to compare GenAI use with other common applications. In decreasing energy use order, he offers these comparisons, which include the energy use in Wh (Watt hours) and water use in litres:
| Energy use | Water use | Type of activity |
| 1000 Wh | 4 L | hour-long Zoom call with 10 people |
| 200 Wh | .8 L | hour-long video streamed on a big TV |
| 30 Wh | 120 cc | generating a page with an online chatbot |
| 20 Wh | 80 cc | charging a smartphone |
| 6 Wh | 24 cc | generating an image online |
| 3 Wh | 12 cc | generating a sentence with an online chatbot |
| .3 Wh | 1 cc | one non-AI Google search |
| .01 Wh | .04 cc | Generating text with a local chatbot [installed on your local computer] |
(table redrawn from the article)
Fairly similar estimates are made in a MIT Technology Review article by James O’Donnell and Casey Crownhart (2025). They also complain about the lack of transparency from companies and suggest that low-scale use is not more troublesome than other things that we may do online, but that the widespread incorporation of AI in everyday software may have a significant multiplier effect. MIT Technology Review has a whole series of articles on AI and energy use, if you are interested in following this up further.
I think the main thing I take away from this is that we aren’t curious enough about our everyday use of digital tools. I think I knew, vaguely, that streaming videos and online meetings are quite heavy users of computer processing power, and thus have an environmental impact, but I didn’t realise how much it was compared to standard internet searches or even GenAI use. I had definitely read in a newspaper somewhere that creating images with GenAI is very resource-intensive, but it doesn’t seem worse than generating large amounts of text, from these figures.
There are plenty of reasons to be cautious about GenAI tools, and we should apply that thinking to all our decisions as users and consumers of digital tools and information.
Note: This is an edited version of some material in our free Canvas course for teachers at Lund University. No GenAI tools were used in writing this post, but I don’t know how much energy my normal Office installation uses.
Author: Rachel Forsyth, Educational developer, Unit for Educational Services, Lund University
Featured image: “Professors see no evil”, generated by Copilot, May 2024
References
Bashir, N., Donti, P., Cuff, J., Sroka, S., Ilic, M., Sze, V., Delimitrou, C., & Olivetti, E. (2024). The Climate and Sustainability Implications of Generative AI.
Ekin, A. (2019). AI can help us fight climate change. But it has an energy problem, too. Horizon: The EU Research and Innovation Magazine.
Ippolito, J. (2025). AI’s impact on energy and water usage v 1.8. University of Maine. Retrieved 07/07/25 from https://ai-impact-risk.com/ai_energy_water_impact.html
O’Donnell, J., & Crownhart, C. (2025). We did the math on AI’s energy footprint. Here’s the story you haven’t heard. MIT Technology Review.
OECD. (2022). Measuring the environmental impacts of artificial intelligence compute and applications. OECD. https://www.oecd-ilibrary.org/content/paper/7babf571-en
