Artificial intelligence and especially its generative variant is popping up in more and more places in our daily existence. The technology promises solutions to all kinds of problems, and we seem to be only at the beginning of a revolution. Many opportunities for application are also being identified for the water sector [1][2], and it is difficult not to submit to the temptation of all kinds of nifty large language model applications. Think of generative AI like a smart autocomplete on your phone. When you start typing a message, it predicts the next word based on what you’ve typed so far. Generative AI and large language models (LLMs) work similarly. They’re trained on massive amounts of text to understand language patterns. When you give them a prompt, they predict and generate text in response, creating coherent and contextually relevant sentences, just like an advanced, imaginative autocomplete! Both in science and in society, its use has normalised. But this comes at a price: a footprint in terms of energy, water and resources.

As voiced by OpenAI frontman Sam Altman, the AI industry itself recognises that the energy demands of future artificial intelligence will be so great that a breakthrough in sustainable energy generation is necessary [3]. But it’s not just about energy; the data centres on which AI runs (among other energy guzzlers like cryptocurrencies and video streaming services) also use large amounts of fresh water for cooling [4]. Furthermore, their production requires large amounts of raw materials, including many metals extracted from mining. Moreover, devices that AI runs on are discarded after a relatively short lifespan [5]. These negative impacts risk being exported, again, to emerging economies in the global south [6].

Several publications in the formal and grey literature in recent years give some insight into the energy, greenhouse gas and water footprints of generative AI. Energy quantities in the range of 300-1,300 MWh are reported for the training phase of well-known previous generation large language models, and 52,000-62,000 MWh for GPT4 [7]. Associated greenhouse gas emissions are in the range of 30-550 up to 15,000 tons CO2eq  [7][8][9][10]. For the deployment phase, estimates are about 3-4 Wh per query; that is about 10 times more than for a traditional Google search [9][10]. For GPT-4, this is yet another factor of 3 higher [11]. To put these numbers in perspective: 1) preparing a cup of tea or charging a smartphone takes about 10-15 Wh, which is thus comparable to 1-3 interactions with a language model; and 2) one of the Dutch water utilities reports a total carbon footprint of just under 10,000 tons CO2eq for 2022 (37 million m³ of drinking water, 300,000 connections). Thus, training GPT-4 has had a larger climate footprint than supplying an entire province with high quality drinking water for an entire year.

Much less is known about the water consumption of data centres than about their energy consumption. Li et al. [12] were the first to estimate this. They conclude that when performing 10-50 inferences with ChatGPT3, 500 ml of water is consumed. This includes water consumed in the data centres and water consumed in electricity generation. Note that both can vary greatly with data centre location and method of electricity generation.

The energy and water demands of generative AI are growing rapidly. By 2027, they are expected to be comparable to those of a country like the Netherlands [12]. Even though several avenues are pursued to reduce the energy hunger of AI under the flag of “green AI”, there is the risk of Jevons’ paradox coming into play – an increase in efficiency that fuels an explosive increase in demand. It has been hypothesized that as a society develops, its environmental footprint first increases and then drops off again – an inverted-U-shaped environmental Kuznets curve – though many emissions seem to follow an N-shaped curve. Wang et al. [13] were unable to draw unequivocal conclusions about the shape of the curve for AI’s environmental footprint. Or in other words, we do not yet know whether the large-scale introduction of “green AI” will ultimately lead to a decrease or an increase in its environmental footprint; whether it will be possible to keep the environmental impact of generative AI within acceptable limits.

At present, however, simple back-of-the-envelope calculations show that it seems unlikely that targeted or even broader adoption of AI by the water industry will result in it becoming a significant factor in its total greenhouse gas emissions. It seems more likely that application of AI will contribute to reducing the industry’s footprint. The real problem is with society-wide implementation and use of AI in a whole range of applications, some useful and some trivial. As a general principle, it is wise to deploy AI where it has positive impact, where it has clear added value in solving a problem, increasing resilience or reducing emissions. And it is wise to avoid using AI for trivial tasks or tasks without clear positive impact. This applies first and foremost to our global society, although it seems to have neither the means nor the self-control to achieve such an ambition; this would require far-sighted and bold political intervention. But the water industry may lead by example.

References

[1] Makropoulos, C., & Savić, D. A. (2019). Urban hydroinformatics: Past, present and future. Water, 11(10), 1959. https://doi.org/10.3390/w11101959

[2] Doorn, N. (2021). Artificial intelligence in the water domain: Opportunities for responsible use. Science of the Total Environment, 755, 142561. https://doi.org/10.1016/j.scitotenv.2020.142561

[3] Reuters (2024) OpenAI CEO Altman says at Davos future AI depends on energy breakthrough. https://www.reuters.com/technology/openai-ceo-altman-says-davos-future-ai-depends-energy-breakthrough-2024-01-16/

[4] Crawford, K. (2024) Generative AI is guzzling water and energy. Nature, vol. 626, doi: https://doi.org/10.1038/d41586-024-00478-x

[5] Hildebrandt, F., & Jung, M. (2023). Digital, grün, global gerecht: Mit Digitalisierung das Klima retten? Forschungsjournal Soziale Bewegungen, 36(3), 392-403.

[6] Mongabay (2024) Critics fear catastrophic energy crisis as AI is outsourced to Latin America.   https://news.mongabay.com/2024/03/critics-fear-catastrophic-energy-crisis-as-ai-is-outsourced-to-latin-america/

[7] Ludvigson, K. (2023) The carbon footprint of GPT-4. https://towardsdatascience.com/the-carbon-footprint-of-gpt-4-d6c676eb21ae

[8] Ludvigson, K. (2023) Environmental Impact of Ubiquitous Generative AI. https://medium.com/p/9e061bac6800

[9] Luccioni, A. S., Viguier, S., & Ligozat, A. L. (2023). Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253), 1-15. https://www.jmlr.org/papers/v24/23-0069.html

[10] De Vries, A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191-2194. https://doi.org/10.1016/j.joule.2023.09.004

[11] Schreiner, M. (2023) Leaks zeigen GPT-4-Architektur, Datensätze, Kosten und mehr. https://the-decoder.de/leaks-zeigen-gpt-4-architektur-datensaetze-kosten-und-mehr/

[12] Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making ai less” thirsty”: Uncovering and addressing the secret water footprint of ai models. arXiv preprint arXiv:2304.03271. https://arxiv.org/pdf/2304.03271

[13] Wang, Q., Li, Y., & Li, R. (2024). Rethinking the environmental Kuznets curve hypothesis across 214 countries: the impacts of 12 economic, institutional, technological, resource, and social factors. Humanities and Social Sciences Communications, 11(1), 1-19. https://doi.org/10.1057/s41599-024-02736-9