As the world meets at the UN Climate Change Conference in Baku, stories about the unsustainability of the data centres that power artificial intelligence are making headlines every day. This column introduces the concept of an ‘eco-political economy of AI’, covering each step from mining materials to processing data to digital waste. Only by examining the complexities of the AI production chain can we begin to understand the true environmental impact of these technologies.
For the last decade, a significant focus within AI studies has revolved around the concern over algorithmic biases, encompassing issues of race and gender discrimination, exclusion, and oppression within AI systems (Eubanks 2018, Noble 2018, Broussard 2023). Facial recognition systems, for instance, exhibit higher error rates for individuals with darker skin tones and demonstrate better accuracy in identifying male or binary genders (Broussard 2023, Noble 2018, Benjamin 2020). Another crucial topic investigated in the field has been the impact of AI on privacy and surveillance (Pasquale 2015). As the world meets at the UN Climate Change Conference in Baku, however, it is stories about the unsustainability of the data centres that power artificial intelligence (AI) that are making headlines every day.
Engineering studies have played a pivotal role in advancing research engaging with the environmental toll of AI and its energy consumption. The pioneering study connecting AI with its environmental costs was published in 2019 by researchers at the College of Information and Computer Sciences at the University of Massachusetts Amherst (Strubell et al. 2019). For the first time, research sought to quantify the energy consumed by running AI programs. In the case examined, a common AI training model in linguistics was found to emit more than 284 tonnes of carbon dioxide equivalent. In more recent times, tools such as the ‘machine learning emissions calculator’ (Lacoste et al. 2019) have become increasingly accessible (Luccioni et al. 2023). Recent studies focusing on ChatGPT have highlighted the urgency of recognising the massive water footprint caused by AI models (George et al. 2023, Heikka 2023).
Although these studies are highly relevant, they often stop short of exploring the complex global ecosystems of technology and the full ecological impact of AI (van Wynsberghe 2022, Kuntsman and Rattle 2019). Since 2017, beginning with a book I co-edited with Graham Murdock (Brevini and Murdock 2017), I developed a research agenda that dives into the ecological harms of AI. Through various contributions (Brevini, 2020, 2021, 2023a, 2023b, 2024), I have introduced the concept of an ‘eco-political economy of AI’ to address the intricate global ecosystems and production chains underlying AI and to understand more comprehensively the complexity of its environmental harms.
The production chain of AI is a series of steps that turn raw materials and data into the services we use. It starts with extracting rare minerals and metals, such as lithium and cobalt, which are used to build the hardware, including powerful computer chips. Then, data are collected, fed into these computers and processed by complex algorithms to ‘train’ AI systems, teaching them how to perform tasks such as translating, understanding speech, or recognising faces. Later, the trained AI is embedded into devices, apps, or services. At this point, users start interacting with, or ‘consuming’, AI. Once these objects reach the end of their life, they need to be disposed of and discarded, becoming ‘e-waste’ or ‘digital rubbish’. Each step – from mining materials to processing data to digital waste – is part of the AI production chain and each step has a considerable environmental toll.
So, the eco-political economy of AI, which addresses the complex environmental costs of the technology, starts with the extraction of rare minerals and metals such as cobalt, copper, nickel, and rare earth elements, which are all crucial for manufacturing AI-powered devices. Demand for these mineral resources is growing at an alarming rate. The European Commission has projected that the demand for lithium in the EU will increase to 18 times the current level by 2030 and to 60 times the current level by 2050 (European Commission 2022).
How sustainable are the practices for the extraction of mineral resources? Besides the political and economic precarity of the supply of these minerals, which is mostly in the Global South (with China being a dominant player in this market), mining operations discharge toxic chemicals and generate vast amounts of hazardous waste, posing serious environmental and health harms (Brevini 2020, Najar 2021). These practices frequently devastate local and indigenous communities and are reminiscent of colonial exploitation, with inhumane working conditions. Take for example the mining of cobalt, often referred to as a ‘conflict mineral’ (Maconachie 2021). Research into the human rights and health concerns of artisanal cobalt mining reveals alarming findings: most miners lack safety equipment, leaving them exposed to cobalt dust, which can lead to asthma, respiratory issues, and even a potentially fatal condition known as ‘hard metal lung disease’ (Maconachie 2021). The extractive nature of this tech-colonialism is evident in the communities that bear the brunt of mining activities, often facing environmental degradation, displacement, and the loss of livelihoods. An example is the Salt Flats of the Atacama Desert in Chile, one of the biggest reserves of lithium in the world. The environmental impact of lithium mining is extensive (Daroqui 2022). Large amounts of freshwater, a crucial resource in this dry area, are diverted from living communities to support lithium extraction processes. Affected local and indigenous groups have criticised the absence of regulatory environmental standards for these projects, which have disrupted indigenous heritage and displaced indigenous communities.
The second stage of the eco-political economy of AI has received more media attention, due to the vast energy demands of data centres: the training of models. The staggering rise in energy and water consumption by data centres driven by generative AI has forced the major ‘digital lords’ (Brevini 2020) to be more open about their energy and water footprints. Following the launch of generative AI services in 2022, both Microsoft and Google reported significant increases: Google’s data centres consumed 20% more water in 2022 compared to 2021 (Google 2023), while Microsoft’s water usage jumped by 34% during the same timeframe (Microsoft 2022). Goldman Sachs predicts that data centres will account for 8% of US energy by 2030, compared with 3% in 2022. Moreover, it is becoming increasingly evident that it is not only the training of AI models that is becoming more energy-intensive but also the ‘inference’ (i.e. the ongoing interactions of users with AI), further escalating the energy demands of data centres (Desislavov et al. 2023).
The last, crucial link in the eco-political economy of AI is disposal. When digital devices are discarded, they become electronic waste, leaving local municipalities with the difficult task of ensuring safe disposal. This task is so onerous that it is frequently offshored, with many countries, primarily in the Global South, becoming digital dumping grounds for more privileged nations. According to the UN’s Global E-waste Monitor (2024), the world’s production of electronic waste is growing five times faster than documented e-waste recycling. A record 62 million tonnes of e-waste was produced in 2022 – up 82% from 2010. At this rate, we are on track to produce 82 million tonnes by 2030. Only 1% of the demand for rare earth elements is fulfilled through e-waste recycling. Generative AI is especially problematic in this regard, as it is driving faster server innovation, particularly in chip design: the latest AI chips such as the Nvidia are driving e-waste up at unprecedented levels.
Only by examining the complexities of the AI production chain can we begin to understand the true environmental impact of these technologies. It is essential to question who should own and control the key infrastructures that drive AI, ensuring that the Climate Emergency is prioritised in these discussions. We must consider for what purposes AI is developed and how it will affect collective wellbeing. What values should steer AI development if we are to effectively address the Climate Emergency?
Various international agreements, position papers, and guidelines are being discussed or initiated both globally and at national levels. Progress is being made, as seen in UNESCO’s recently adopted recommendation on artificial intelligence, which states that AI systems should not be used if they have a disproportionate negative impact on the environment (UNESCO 2021). Unfortunately, the recently adopted EU AI Act, with its ambitious global regulatory drive, is a missed opportunity as it fails to include clear strong environmental regulations imposed on AI providers or deployers. Under the AI Act, providers have the option to develop and implement codes of conduct, which may incorporate voluntary commitments to environmental sustainability.
Adopted in March this year, the new climate disclosure rules from the US Securities and Exchange Commission (SEC 2024), which require companies to report their emissions as they are deemed financially material to investors, are one step in the right direction. These types of interventions could be coupled with an AI carbon footprint label that not would only provide details of the energy used for training, but also clearly list details about the raw materials used, the overheads, recycling options, and disposal costs. This would help increase public awareness of the environmental implications of adopting a particular AI-powered device. Bringing transparency to the making of energy used in producing, transporting, assembling, and delivering everyday AI technology would empower policymakers to make better-informed decisions and enable the public to make more conscious choices. Global policymaking should also promote educational programmes that enhance green AI literacy, raise awareness of the environmental costs of AI, and emphasise the importance of responsible energy consumption. These programmes could include initiatives to ban the production of overly data-intensive and energy-depleting products.
In March 2023, the Intergovernmental Panel on Climate Change (IPCC), made up of the world’s leading climate scientists, released its last report. This was dubbed the ‘final warning’ to limit global temperature rises to 1.5°C above pre-industrial levels, the threshold beyond which our damage to the climate will become irreversible (IPCC 2022). To keep within the 1.5°C limit, emissions need to be reduced by at least 43% by 2030 compared to 2019 levels, and at least 60% by 2035. Considering the ever-increasing environmental toll of AI development, these targets seem to be slipping further out of reach. The responsibility to prioritise the Climate Emergency in AI development falls to us. We cannot afford to delay this any longer.
Source : VOXeu
