The environmental costs of AI
AI is an important topic and it’s driving a new revolution that, like the industrial and information revolutions, will transform how we live, work and impact the environment.
The environmental impacts of our information and communication technology has long been an under-recognised issue.
Starting with web and mail servers in the 1990s, there has been a steady growth in data centres, infrastructure and energy use. The rise of cloud computing, cryptocurrency mining and on-demand streaming have been major growth factors. Since ChatGPT burst onto the scene in 2022, we’ve been in the middle of a step-change in our use of computing resources and associated environmental impacts.
Using ChatGPT to generate text uses somewhere between 10 to 90 times more energy per query than a conventional Google search.
Producing an image via generative AI uses about 20 times more energy than a generative AI text query.
This year, producing videos via generative AI is taking off, and the energy to produce videos is likely be an order of magnitude higher again.
These AI models are now run millions, or possibly billions, of times every day. Their popularity is growing, and they are being embedded within many of the software platforms we all regularly use.
Underpinning all this energy use are the data centres that supply the computing resources to run generative AI models.
They currently use between 1-3% of the world’s energy, and this is set grow.
Australia has a significant number of data centres that use approximately 5% of Australia’s electricity supply.
Increasing the efficiency of data centres is one important policy response, however this alone won’t solve the problem.
One such problem is that large generative AI models can be based in data centres anywhere in the world, and most of Australia’s AI use is ‘offshore’.
Another important factor in reducing environmental impacts could be encouraging greater care in the choice of AI models for different applications; smaller, more specialised AI models have been shown to be vastly more energy efficient than larger general-purpose models.
Appropriately managing the environmental footprint of AI is going to be difficult, and increasing efficiency of data centres and AI models is unlikely to make much difference in the short term due to the rapid increases in AI use we’re currently seeing.
It is critical we increase the public awareness of the environmental footprint of AI, and more generally of the software platforms we regularly use.
Ascelin Gordon is a Senior Lecturer in Sustainability and Urban Planning at RMIT. He develops quantitative and qualitative approaches for understanding the impacts of environmental policies using machine learning to interrogate large spatial datasets.
