Artificial intelligence has been discussed at previous climate conferences: as early as 2021 in Glasgow, there were discussions about both the benefits of AI, and the probable need for additional resources to power it.
Artificial intelligence has been discussed at previous climate conferences: as early as 2021 in Glasgow, there were discussions about both the benefits of AI, and the probable need for additional resources to power it[1]. More recently, it even made a minor appearance in the first global stocktake, where it was mentioned as a potential “tool for advancing transformative […] solutions”[2]. Despite this, two things are likely to change in the way it is discussed at the next COP and those that will follow.
First, the impact of AI on potential climate solutions is now more concrete than ever. From predicting demand in energy systems to powering innovation in industries with hard-to-abate emissions, it is now easier to see how transformational AI could be. It has already improved the professional lives of many and is enthusiastically embraced by companies in search of higher productivity.
Second, that very enthusiasm has led many to worry about our ability to satisfy the increased demand for electricity that a generalised use of AI would imply. Whilst we can hope the scaling up will be done in a low-carbon fashion, this should not be a given. Current energy systems already require large investments to be decarbonised. Attaining the same objective while satisfying a significant amount of additional demand will be more difficult, and something that was certainly not on the cards when the Paris Agreement was signed at COP21.
The technology sector represents only a few percentage points of global emissions[3], but the current explosive growth could lead to much higher levels. The exact level is beyond the scope of this article but will surely be hotly debated. The problem is large enough that most of the big tech companies will probably need to redraw their climate plans. They were mostly enthusiastic champions of the climate cause early on. In 2020, Microsoft announced an impressive plan to aggressively reduce its own emissions (including Scope 3 emissions) while investing in carbon-removal technologies. Four years on, overall emissions are significantly higher, largely due to the increased power needs of servers and other infrastructure that has come out of the AI boom. Microsoft is sticking to its target for now, but the clock is ticking, and it looks like the whole tech sector will face similar conundrums. The fact that these companies now appear eager to sign agreements with nuclear energy producers further underscores their urgent need for additional energy[4][5].
Of course, efficiency gains will be made along the way: new AI chips are expected to be many times more energy-efficient than the current ones and some of them will help us to reduce emissions in other ways. But the Jevons Paradox comes to mind[6]: will this increased efficiency be used to reduce emissions or to increase usage? There is a serious doubt that it will be the former. The track record of modern economies is very poor in this field, and the promise of AI-enabled toothbrushes[7] does not fill one with hope.
Nevertheless, doubling energy efficiency by 2030 remains a prominent goal in the global stocktake, and something that a lot of companies and investors are working towards. According to the International Energy Agency, doubling energy efficiency is one of the three key targets that could deliver most of the effort needed for a 1.5°C scenario by 2030, alongside the tripling of renewable capacity and a drastic reduction in methane emissions[8]. But to make efficiency work in favour of climate, we will need to decide what counts as efficiency gains (and hence a potential solution) and what should be seen as an increased energy need sending us off track. So far, there are no signs of this debate in the current negotiations. If private companies must redraw their climate plans, soon countries will have to do the same, and if so, how will their Nationally Determined Contributions be impacted? Would it be acceptable for a country to diverge from its existing commitments due to AI emissions in the hope that this same technology will help to reduce emissions in the future? If so, what would be an acceptable time horizon?
These questions need to be answered, and as is often the case when climate is concerned, it needs to be both grounded in science and politically acceptable. For those reasons, I hope AI, with all its implications, will be hotly debated at COP29 and have a significant presence in the final agreement.
[1] “Climate change and AI: Recommendations for Government Action”, Global Partnership on AI Report, November 2021, https://www.gpai.ai/projects/climate-change-and-ai.pdf
[2] “Outcome of the first global stocktake”, https://unfccc.int/news/cop28-agreement-signals-beginning-of-the-end-of-the-fossil-fuel-era
[3] The Carbon Emissions of Big Tech, Rodrigo Navarro, https://www.electronicshub.org/the-carbon-emissions-of-big-tech/
[4] Amazon and Google Sign Nuclear Energy Deals as AI Power Demands Surge, https://www.cnet.com/tech/services-and-software/amazon-and-google-sign-nuclear-energy-deals-as-ai-power-demands-surge/
[5] Why Microsoft made a deal to help restart Three Mile Island, https://www.technologyreview.com/2024/09/26/1104516/three-mile-island-microsoft/
[6] https://en.wikipedia.org/wiki/Jevons_paradox
[7] “80% of people miss at least one area of their mouth when brushing” says one advert about a toothbrush “powered by Artificial Intelligence [which] tracks where your brush [is] and gives you real-time feedback for your best results every day.”
[8] Net Zero Roadmap: A Global Pathway to Keep the 1.5 °C Goal in Reach, https://iea.blob.core.windows.net/assets/9a698da4-4002-4e53-8ef3-631d8971bf84/NetZeroRoadmap_AGlobalPathwaytoKeepthe1.5CGoalinReach-2023Update.pdf
