AI’s Dirty Secret

Introduction

Artificial intelligence (AI) is transforming industries, reshaping economies, and redefining daily life. Whether you sit on the pro- or anti-side of the debate (or don’t feel particularly strongly either way), there’s no denying the prevalence of AI technology. Ever since ChatGPT burst into the public consciousness in 2022, AI has featured increasingly prominently in societal debates, from whether it might innocently assist us with tiresome administrative tasks to whether it is going to wipe out the entire human workforce and perhaps accelerate the speed of our species’ destruction to boot.

Those on the pro-side of the debate point to the many innovative possibilities AI might afford us, most especially in science and technology, with further breakthroughs in sectors like healthcare that are objectively exciting. Those on the anti-side tend to focus more on the human impact –– what might wide-scale job displacement do to our society going forward? And yet, one further element is often excluded from the debate: that of artificial intelligence’s ecological impact.

Behind the seamless digital experiences powered by AI lies a hidden truth: the technology is an energy-intensive beast. Its appetite for electricity, water, and rare earth metals is accelerating, creating environmental consequences that threaten to overshadow its advancements. This article explores AI’s environmental footprint, its implications, and the actions needed to mitigate its impact.

An energy goliath

A generative AI system may use 33 times more energy to complete a task than it would take with traditional software [1]. Meanwhile, one estimate posits that the amount of computational power used for AI is doubling roughly every 100 days [2]. Given there are more than 100 million users of ChatGPT every week –– not to mention the number of users on other platforms –– it’s not hard to see why energy use is skyrocketing [3].

A large amount of the ecological harm comes from the training process. Recent research shows that training GPT-3 consumed approximately 1,287 megawatt-hours (MWh) of electricity, emitting 502 metric tonnes of CO2, which is roughly equivalent to the emissions of 112 gasoline-powered cars over a year [4]. This is only the training phase. The power required for “inference” — when models process real-time queries — can account for up to 60% of AI’s total energy consumption [5]. A separate study by researchers at the University of Massachusetts showed that training a single large AI model could generate a carbon footprint of 626,000 pounds of CO2 — equivalent to five times the lifetime emissions of one car. [6]

Unsurprisingly, then, the companies responsible for pushing the AI revolution have put their sustainability credentials under pressure. Between 2020 and 2023, Microsoft’s disclosed annual emissions increased by around 40%, from the equivalent of 12.2 million tonnes of CO2 to 17.1 million tonnes [7]. Meta disclosed in 2023 that its Scope 3 emissions had increased by over 65% in just two years, from the equivalent of 5 million tonnes of CO2 in 2020 to 8.4 million tonnes in 2022. [8]

Then there’s Google, whose emissions were almost 50% higher in 2023 than in 2019 [9]. The company’s 2024 Environmental Report also reveals that its overall GHG emissions have steepled by 13% in a year since that 50% rise [10]. Speaking in the introduction to the report, Google CSO Kate Brandt and Benedict Gomes, SVP, Learning & Sustainability, said: “While we remain optimistic about AI’s potential to drive positive change, we’re also clear-eyed about its potential environmental impact and the collaborative effort required to navigate this evolving landscape.”

The data centre dilemma

AI relies on data centres to function, with these facilities housing the servers and GPUs needed to train and operate AI models. In 2023, before the AI boom really kicked off, the International Energy Agency estimated data centres already accounted for 1–1.5% of global electricity use and around 1% of the world’s energy-related CO2 emissions [11]. This figure is climbing as AI adoption expands.

A large part of the problem is that data centres generate a lot of heat and consume large amounts of water to cool their servers. According to a 2021 study, data centres in the United States use about 7,100 litres of water for each megawatt-hour of energy they consume [12]. Google’s US data centres alone consumed an estimated 12.7 billion litres of fresh water in 2021 [13]. The International Energy Agency forecasts that by 2030, AI energy consumption will make up 20% of global electricity supply if current growth trends continue. [14]

Rare earth metals and E-waste

A further issue is that the hardware powering AI — such as GPUs, CPUs, and specialised chips — relies on rare earth metals like lithium, cobalt, and nickel. Extracting these materials is energy-intensive and environmentally harmful. Mining operations contribute to deforestation, soil degradation, and significant carbon emissions. They also exploit labour, particularly in developing countries like the Democratic Republic of Congo, where cobalt mining is rife with human rights abuses.

Electronic waste (E-waste) adds another dimension to the problem. As AI-driven hardware is rapidly replaced by more advanced systems, obsolete equipment contributes to a global e-waste crisis. E-waste contains dangerous chemicals that contaminate the environment when discarded. The World Economic Forum (WEF) already projects that by 2050, generated e-waste will have surpassed 120 million metric tonnes [15]. That’s the equivalent of nearly 12,000 Eiffel Towers of waste. The increasing demand for natural resources like water and earth metals to power AI hardware is set to prove ethically divisive, as rich, tech-driven countries mine less economically developed, resource-rich countries who are both less likely to feel the benefits of AI and more likely to suffer from the environmental impacts of climate change.

Greenwashing or genuine solution?

To counterbalance its environmental costs, AI is increasingly being marketed as a tool for sustainability. Companies tout its ability to optimise energy grids, improve efficiency in logistics, and assist in climate modelling. Some reports predict that AI has the potential to help mitigate 5-10% of global GHG emissions by 2030. [16]

Google’s report said the company was “advancing climate action through AI in three key areas” [17]:

Organising information: Fuel-efficient routing uses AI to analyse traffic, terrain and a vehicle’s engine to suggest the most efficient route. It’s estimated to have helped enable more than 2.9 million metric tonnes of GHG emissions reductions since the feature launched in late 2021 to the end of 2023
Improving prediction: Google said it built a breakthrough global hydrological AI model and combined it with publicly available data sources to predict floods up to seven days in advance in over 80 countries
Better optimisation: Green Light is an AI-based tool that helps city traffic engineers optimise the timing of traffic lights to reduce stop-and-go traffic and fuel consumption

What should be noted is that these are claims the company is making about itself. It can (and has) been accused of marking its own homework when it comes to the sustainable offering it provides regarding AI, with some allegations of greenwashing (i.e. highlighting AI’s potential benefits while downplaying its systemic environmental challenges.) For every application that reduces emissions, countless others, such as cryptocurrency mining or generative AI art, exacerbate environmental harm. How far the scales tip in one direction or the other is yet to be determined.

The future

The growth trajectory of AI is exponential. A 2024 report by the International Energy Agency warns that, if unchecked, AI’s energy consumption could double by 2026, equivalent to the annual electricity usage of Japan [18]. As AI models grow larger and more complex, they will require even more powerful hardware and data infrastructure, exacerbating the resource strain.

Moreover, the increasing reliance on AI in critical sectors — such as healthcare, finance, and autonomous vehicles — means that scaling down its usage is not a feasible option. Instead, solutions must focus on reducing its environmental impact without stifling innovation.

Solutions

The major players in the AI space are aware that the negative environmental impact of AI could be harmful to their brand image (and the planet, though we’ll let you decide which they prioritise). As such, and as Google’s above claims make clear, they are searching for solutions –– ways to make AI less severe in its energy consumption. For example, researchers are designing specialised hardware such as new accelerators, new technologies such as 3D chips, which offer much-improved performance, and new chip cooling techniques. Computer chip maker Nvidia claims its new ‘superchip’ can deliver a 30 times performance improvement when running generative AI services, while using 25 times less energy. [19]

Quantisation is also touted as an enhancement to the existing system. It reduces the numerical precision of AI calculations — which raises issues among developers — but it leads to as much as 50% computational cost savings, helping AI systems scale down computational costs to more manageable levels and reduce energy consumption [20]. Then there are the emerging technologies like neuromorphic computing, which mimics the human brain’s neural structure and uses 1,000 times less energy than traditional CPUs. [21]

AI’s dirty secret

While artificial intelligence is undeniably a powerful force for innovation and progress, its environmental costs demand urgent attention. From the staggering energy consumption and water use of data centres to the ecological devastation caused by mining rare earth metals, AI’s environmental footprint poses a significant challenge. Despite its potential to contribute to sustainability through applications in energy optimisation and climate modelling, the current trajectory of AI development risks exacerbating the climate crisis unless swift and meaningful action is taken.

The solutions exist — whether through technological advancements like more efficient chips, innovative cooling systems, and neuromorphic computing, or policy measures that incentivise greener AI practices. However, the balance between fostering innovation and mitigating harm requires a coordinated global effort. As we navigate this rapidly evolving landscape, the challenge is not just to reduce AI’s environmental impact but to ensure that its promise as a transformative tool is not overshadowed by its hidden costs. Only then can we truly harness AI as a force for good without leaving a trail of ecological destruction in its wake.