Optimistic advocates for AI say this environmentally costly technology will become more sustainable with time. We can use AI more efficiently, and explore less energy-intensive designs inspired by the human brain. We can build data centres more sustainably, using wood or low-carbon concrete and steel. The heat from the data centres can warm homes in the local area.
Of course, if we start using AI systems too widely (including where we don’t really need them), the growth might outweigh any of these potential gains in efficiency. But recently, I’ve been hearing another argument: that AI itself is tackling climate change. AI can help to model wildfires, optimise energy consumption to stabilise the grid, accelerate the development of low-carbon materials, and much more.
My research team recently published a report that digs into these claims – and found some cause for concern.
As it turns out, it’s difficult to compare the environmental impacts of asking AI to carry out a task versus asking people. Take writing and making art. One study says that it is much more environmentally friendly for AI to do these creative tasks than humans.
But one of its methods for investigating this is to allocate a slice of the carbon footprint from all human activities (eating cows, catching planes and so on) to the creative activity. By this logic, there would be the same carbon emissions from an hour’s work by an artist, a dairy farmer, a billionaire CEO, or an ecologist restoring a wetland.
A related paper acknowledges the simplification and instead allocates carbon just for the electricity a human consumes. The authors say this is a practical approach to carbon accounting. Establishing a fair comparison between human and AI work is hard because the tasks may look alike, but their underlying processes are fundamentally different.
What about the current total carbon impact of AI? Another paper, partly funded by Microsoft, mentioned that AI is today responsible for just 0.01% of global carbon emissions.
When we looked closer we found this figure was based on the emissions of one year’s AI server sales by Nvidia – the biggest manufacturer of such hardware – as estimated in one analysis. This prediction hasn’t been verified, and if it is accurate, it wouldn’t include AI being run on servers installed in previous years.
When contacted for comment, some of the authors said that this estimate wasn’t the article’s main focus. Assessing the climate impacts of AI is complicated because we don’t know how future AI models will be built, operated and used, they added.
Another study reviewing AI’s sustainability benefits cited several other articles that, in our view, appeared to feature mistakes – like referring to studies to back up their claims which did not contain relevant information. AI systems sometimes make mistakes – known as “hallucinations” – like when Microsoft Copilot accused a journalist of committing the crimes he had reported on.
So we asked the authors if they had used an AI to write the article. They dismissed this idea and stood by the integrity of their review. They agreed that not all of the AI sustainability solutions their review referred to can be supported by existing studies. But they said that they had also included predictions about what AI might one day be capable of, based on their own expert judgement, which they said was standard practice.
The authors also pointed out that their work has been widely shared without complaint. But the fact that articles can be read widely without eliciting alarm is exactly the issue. Would every reader assume that the article was predicting what AI might one day do, rather than explaining what AI can already do?
Overall, how reliable is current research on AI and sustainability? We’re not sure. We haven’t yet had the chance to conduct a more comprehensive investigation. But assessing AI’s future potential requires a clear understanding of its achievements to date. Such an investigation is urgently needed.
Meanwhile, we should avoid lumping all kinds of AI together.
There are in fact diverse forms of AI: big, small, discriminative, generative, machine learning, symbolic and more. I can be excited about an AI that excels at counting carrots, and helps farmers to plant them more effectively, without offering a blanket endorsement of all AI systems.
There are different types of climate action too. Climate mitigation is about getting carbon emissions down to net zero to stop global warming. Climate adaptation is about learning to live and thrive in a warmer world. We need both.
AI for climate adaptation is very welcome indeed (say, helping us to increase carrot yields, despite more volatile weather). But it doesn’t simply offset the carbon cost of AI. It would be like comparing apples with oranges (or carrots). It’s a tricky calculation to make, one with political and ethical dimensions.
Whenever AI is celebrated for amazing achievements, let’s remember that it had some human help. What time, effort, energy and other resources were invested in the project? Could similar results have been achieved using more traditional data collection and analysis, potentially at a lower environmental cost?
The authors of our report have mixed views on AI. None of us are against it, or against using it to solve environmental problems. But to properly govern AI’s net impact on the climate, its benefits must not be overstated.
Don’t have time to read about climate change as much as you’d like?
Get a weekly roundup in your inbox instead. Every Wednesday, The Conversation’s environment editor writes Imagine, a short email that goes a little deeper into just one climate issue. Join the 40,000+ readers who’ve subscribed so far.
Jo Lindsay Walton has received funding from Innovate UK and the Arts and Humanities Research Council for research relating to this article.
