AI’s environmental impact is a live debate
Lately I have run into several discussions about the environmental impact of AI, and of language models such as ChatGPT in particular. The common worry often boils down to one question: how much more energy does a single ChatGPT query use compared with, say, an ordinary Google search?
Examining the environmental impact of AI use is extremely important, but it is equally important in that debate to put those impacts in proportion to the other digital things we do every day.
How much energy does AI really use?
Average estimates for the energy consumption of different digital services look like this:
One ChatGPT query: about 3 Wh
One Google search: about 0.3 Wh
One hour of Netflix: about 900 Wh (equivalent to roughly 300 ChatGPT queries)
It is worth noting that video streaming or Netflix use rarely raises the same concern about energy consumption as AI does, even though in many use cases their consumption is significantly higher.
Language models’ water use – a lot or a little?
Language models and the data centres they require consume a great deal of water, particularly for cooling. That consumption is worth putting in proportion to other everyday water use:
20–50 ChatGPT queries: about 500 ml of water (one water bottle)
A 1-hour Zoom meeting: about 1,720 ml of water (equivalent to about 57 ChatGPT queries)
Watching 10 minutes of 4K video: about 2,580 ml of water (equivalent to about 86 ChatGPT queries)
Producing one hamburger: about 2,400 litres of water (equivalent to as many as 84,000 ChatGPT queries)
Favour a holistic approach
When AI’s environmental impact comes up, it is easy to focus on individual actions, when a more holistic approach would often be more useful. I have said it before and I will say it again: the most important question is not whether AI should be used, but how and for what it is worth using.
Light office work, for example, uses about 20–30 W per hour above basal metabolism. If AI lets us cut our daily working time by, say, two hours, the energy saving could be as much as 40–60 Wh a day. On that basis, 13–20 ChatGPT queries a day could be a more energy-efficient solution than working the same things out manually.
In interpreting the data, distinguish direct from indirect water use
It is also worth noting that calculations of ChatGPT queries’ energy and water use include both the AI models’ training phase and the individual queries themselves. Data centres’ direct water use, however, accounts for only about 15% of the total. Looking at direct data-centre consumption alone, a one-hour Zoom meeting could correspond to as many as 1,030 ChatGPT queries.
In assessing environmental impacts, then, it is critical to put AI use in proportion to the other digital services of everyday life and to find the right balance between the benefits and the environmental load.
Key sources used in this article
U.S. Energy Information Administration (EIA) – energy consumption data: 🔗 www.eia.gov
Electric Power Research Institute (EPRI) – research on energy consumption: 🔗 www.epri.com
Founders Pledge – Climate & Lifestyle Report – analysis of individual climate impact: 🔗 www.founderspledge.com/research/climate-and-lifestyle-report
Sunbird DCIM (Water Usage Effectiveness – WUE) – data centre water use: 🔗 www.sunbirddcim.com
GovTech – estimate of ChatGPT’s water use: 🔗 www.govtech.com
ArXiv.org – academic preprint archive, including research on AI energy consumption: 🔗 www.arxiv.org
Industry Leaders Magazine – Netflix’s energy consumption: 🔗 www.industryleadersmagazine.com
International Telecommunication Union (ITU) – the internet’s energy consumption: 🔗 www.itu.int
Thunder Said Energy – the internet’s total energy consumption: 🔗 www.thundersaidenergy.com
Enjoyed this article? Subscribe to get notified about new Tekoälyfoorumi articles straight to your inbox. We’d also hugely appreciate it if you left us a review on Google. Thank you. 💙
