Artificial intelligence (AI) is often perceived as a driver of innovation and competitiveness, and hundreds of billions of euros in investments have recently been announced in Europe and the United States. In the context of the climate emergency, many voices argue that AI could optimize numerous processes and help reduce the environmental impact of our lifestyles. Others, on the contrary, point to the enormous negative impacts of AI. While digital technology already accounted for 1.8% to 3.9% of global emissions in 2020, the massive construction of new data centers for AI is seen as an additional source of environmental destruction.
So, is AI an ally or a threat to the environment?
The answer is neither simple nor unequivocal: it depends heavily on how AI is used and on the types of AI involved. We will look at how AI can support the ecological transition, but also at the significant impact of this technology, particularly in terms of greenhouse gas emissions and consumption of water, mineral resources, and energy.
1. AI for Green: what can AI do for the environment?
1.1. Two types of AI for the environment
AI applications that benefit the environment can, broadly speaking, be divided into two major categories. The first includes applications dedicated to optimizing physical systems: they aim to reduce waste, improve energy efficiency, or adjust resource consumption in real time. The second concerns the use of AI for the production or sharing of knowledge, whether to model complex phenomena such as the climate or to make scientific information accessible to a wider audience. For example, AI models are used to refine climate simulations, to generate summaries that facilitate the popularization of environmental content, or to automatically detect illegal environmental destruction.
In this first section, we will focus on optimization applications. These are the applications that interact directly with infrastructure and physical systems and can therefore have an immediate and measurable impact on resource consumption and greenhouse gas emissions. This does not mean that AI applications for the production and sharing of knowledge are secondary: their impact on public debate, although difficult to quantify, could prove just as decisive in the long term (one resource among the many available on the subject).
1.2. AI for optimization
One of the most emblematic areas where AI-driven optimization can contribute is data centers. For example, in Google's data centers, AI reportedly makes it possible, according to an article published in 2016, to adjust cooling systems in real time based on dozens of parameters such as server load and outdoor temperature. This system reportedly resulted in an overall reduction of around 15% in the data center's energy consumption. These fine-grained adjustments, made possible by continuous data analysis, can therefore lead to significant resource savings in this particular case.
This example illustrates a broader trend: artificial intelligence could make production processes more efficient across the board. By optimizing resource use, AI can reduce some of the waste that exists today. In the field of energy production, for example, the International Energy Agency (IEA) estimates in its AI and Energy report (2025) the effects of adopting AI-driven systems. The agency estimates that widespread adoption of AI in the electricity sector could generate up to USD 110 billion in annual savings by 2035. These savings would come in particular from better power plant management, predictive maintenance, and smoother integration of renewable energy into electricity grids. The same report estimates that light industries such as electronics and machinery manufacturing could reduce their energy consumption by 8% by 2035 through AI-driven optimization of production flows.
1.3. A limited positive impact
These projections from the IEA, a leading institution in energy forecasting, confirm that artificial intelligence can play a significant role in improving energy efficiency. AI can also, through similar optimization mechanisms, help limit other negative impacts of human activities, such as water consumption or the release of pollutants into the environment.
However, these figures also show that these potential gains, although significant, remain too limited. The 8% of energy that could potentially be saved in light industries must be put into perspective with the radical changes required to decarbonize our lifestyles. For example, France's National Low-Carbon Strategy (2nd version) sets a target of an 81% reduction in industrial emissions between 2015 and 2050 in order to achieve carbon neutrality. It is therefore clear that AI deployed on its own is not enough, and can only be considered part of the solution. Claims that AI could solve climate change, promoted by some major figures in the tech industry, are therefore currently unfounded.
1.4.The importance of indirect effects
We have discussed the direct effects of AI that could be beneficial. However, these will also be accompanied by indirect effects, which could prove even more significant. One of the best-known phenomena in this area is the rebound effect, or more specifically, in its most severe forms, what is known as the Jevons paradox.
This paradox was formulated by economist William Stanley Jevons. He observed that improvements in the efficiency of the steam engine did not lead to a reduction in coal consumption, as one might have expected. On the contrary, this technological improvement made the use of steam more economical and therefore more attractive to industrialists. As a result, the use of steam engines became widespread, and total coal consumption increased sharply. This paradox highlights a fundamental mechanism of energy economics: increased technological efficiency can lead to the expansion of that technology, cancelling out or even exceeding the expected savings.
The same reasoning applies perfectly to the efficiency gains enabled by AI. For example, if AI makes it possible to reduce the energy consumption of an industrial or transportation process, it is plausible that this would reduce operating costs, increase profitability, and therefore encourage greater use or expansion into other areas. This phenomenon is a major counterargument to optimistic projections about AI's environmental potential. (reference article on the subject).
Therefore, while AI can be a powerful optimization lever, it will not automatically reduce our ecological footprint. And even before considering its significant negative impacts, it is essential to recognize that AI is at best one lever among many for achieving a sustainable way of life.
2. The impact of AI: not all AI is created equal
Before looking in detail at the environmental impacts of AI, it is essential to remember that there is no single form of AI, but rather many different forms of artificial intelligence, with varying environmental implications. In particular, it is important to distinguish between traditional machine learning and generative AI.
Machine learning has existed since the 1980s and has developed considerably since then. It encompasses numerous statistical and computational methods (regression, decision trees, neural networks, etc.) that make it possible to identify patterns in data, make predictions, or take automated decisions based on learning from that data. These models are now ubiquitous in digital services: recommendation algorithms, search engines, spam detection systems, or industrial optimization tools. Their effectiveness often relies on large volumes of data, but their architectures remain relatively lightweight, and their energy costs, although real, are generally manageable, particularly once the models have been trained.
Generative artificial intelligence is a subcategory of machine learning based on specific neural networks capable of producing original content: text, images, audio, video, or even code. These models no longer simply predict or classify based on data; they generate new content based on training with massive quantities of data. The turning point came in 2017–2018 with the emergence of models such as GPT and BERT in academia. However, it was mainly from 2022 onwards that their use became widespread among the general public, with applications such as ChatGPT, DALL·E, and Midjourney.
Generative AI developed later because it requires vastly greater amounts of data and computing power than traditional models. Their training requires dedicated supercomputers to operate for weeks, or even months, and large-scale use also generates continuous energy consumption, particularly when these systems are integrated into consumer-facing interfaces such as ChatGPT. As we will see, this new family of AI is highly resource-intensive, raising questions about its compatibility with a responsible approach to digital technology.
3. The impact of generative AI
3.1. The gap between generative and “traditional” AI
To properly understand the difference between the various AI families, it is useful to compare the orders of magnitude of the energy required for their training. Take DINOv2, for example, a non-generative computer vision model published by Meta in 2023. The graphics cards used to train this model consumed approximately 8.8 MWh, enough to power an average French household for around two years.
Computer vision is among the most data- and energy-intensive non-generative tasks. Yet training DINOv2 consumed 2,000 times less energy than training a current language model. Training LLaMA 3.1, a GPT-4 equivalent developed by Meta in 2024, required approximately 21 GWh to power the graphics cards used for training (roughly the amount of electricity produced by a nuclear reactor in one day).
This figure applies only to the graphics cards. The consumption of other IT equipment: processors, motherboards, routers, storage, etc., would also need to be added, as would data center consumption: cooling, lighting, etc. as well as the energy consumed to manufacture this equipment.
3.2. The carbon impact of ChatGPT
This explosion in computing power requirements for generative models does not only concern the training phase. In reality, the use of these AI systems — inference — may eventually consume as much as, or even more than, training, particularly when they are used at very large scale. For example, ChatGPT currently has more than 400 million weekly users, generating billions of interactions every month. Each of these requests requires several graphics cards in a data center, sometimes for several seconds.
According to an estimate published by the French collective Écologits, a simple email generated with ChatGPT would consume approximately 14.9 Wh, or nearly 50 times more than a Google search, which was estimated at 0.3 Wh in 2009. Even though these figures are only estimates (due to the lack of transparency from companies performing inference), the order of magnitude is nevertheless meaningful: a generative model probably consumes significantly more energy than a traditional search engine.
3.3. Other direct impacts
The impacts are not limited to electricity. Intensive inference by generative AI also results in high water consumption, particularly for server cooling. One study estimated that approximately 500 mL of water is evaporated for every 10 to 50 medium-sized requests made using GPT-3. This consumption, sometimes occurring in regions already affected by water stress, represents a major environmental issue at the local level.
Finally, the material footprint of these technologies is considerable. The manufacturing, transportation, maintenance, and end-of-life treatment of the infrastructure required (servers, graphics cards, storage systems, networks, etc.) generate significant impacts. Unfortunately, all AI players maintain a very high level of opacity regarding their operations, and hardware manufacturers are no exception. As a result, there is little information available about the life-cycle impact of AI-specific graphics cards.
The sustainability reports of major companies can be used to obtain an initial estimate. Meta and Google estimate, for example, that their data centers generate more CO₂ through Scope 3 emissions (indirect emissions) than through Scope 2 emissions (emissions related to energy consumption). In other words, if the complete life cycle of hardware is assessed, the energy used to operate it represents less than 50% of the CO₂ emitted.
We have seen that generative AI consumes a great deal of energy, but this observation reminds us that direct consumption is the most visible impact, but not necessarily the most significant. It is therefore essential to consider as broad a range of indicators as possible in order to obtain a complete picture of AI's environmental footprint.
3.4. Beyond water and electricity
Analyzing CO₂ emissions and water consumption is the easiest approach, but it should be complemented by consideration of many other impacts: soil and groundwater pollution, damage to biodiversity, competing uses of resources, depletion of scarce resources, etc. There are currently no precise figures for these impacts, and the very high level of opacity across the sector makes it difficult to build knowledge on the subject.
More broadly, the environmental footprint of generative AI is only one part of its still poorly understood overall impact. A discussion of the social impacts of AI is beyond the scope of this article, but we emphasize that it is essential to consider AI by taking all types of impact into account.
Approaches that focus only on certain parts of the problem (only CO₂, only environmental impact, etc.) carry a risk. They can lead to solutions that improve the indicators being studied while creating new negative impacts in areas that have not been analyzed. For example, banning a digital service for environmental reasons cannot be done without considering the social and economic impact of that service. We therefore encourage readers to explore the many resources available on the other impacts of AI.
Regarding the environmental footprint, the following conclusion can be drawn: generative AI has a greater impact than most other digital services because of its increased computing requirements, which result in significant consumption of hardware, energy, and cooling resources.
3.5. Explosive adoption
Despite these multiple impacts, generative AI is growing rapidly, and numerous data centers are currently under construction or being planned. In April 2025, the IEA predicted that global data center energy consumption could double by 2030 compared with 2024, and triple compared with 2020. This increase is primarily driven by the growth of accelerated servers, specifically designed for generative AI.
While digital technology already accounts for several percent of the global carbon footprint, this significant growth and the impacts we have described are clearly at odds with environmental preservation objectives. It is therefore becoming urgent to integrate the real cost of these technologies into discussions about their development.
Source: IEA — Energy and AI (2025)
4. The importance of an impact-inclusive approach
We have seen the potential of certain “green” AI applications, as well as the environmental consequences of widespread AI adoption (and particularly generative AI). So ultimately, are we for or against AI? Will technology save the planet or spell the end of the Paris Agreement? Rather than making a binary judgment, it is essential to adopt a nuanced perspective.
Artificial intelligence, in all its forms, is above all a technology of acceleration. It makes it possible to automate tasks, optimize processes, and generate results at unprecedented speed and scale. This ability to accelerate, which is inherent to all AI systems, whether traditional machine learning or generative models, lies at the heart of both its potential for environmental transformation and its risks.
When applied to clear objectives of resource efficiency and sufficiency (energy optimization, waste reduction, better allocation of resources), AI, particularly in its traditional form, can be a valuable technical ally in reducing the environmental impact of existing systems. As we have seen, these technologies can be relatively lightweight, and their targeted use can deliver significant gains in industrial and digital infrastructure.
Conversely, when AI is used to accelerate the production of content, interactions, or digital services without careful consideration, as is often the case with generative AI, it can become a driver of intensification of environmental pressures. The automation of tasks that were previously manual or limited can lead to rapid growth in usage and therefore to proportional increases in emissions, water consumption, and material requirements.
In this context, the goal is not to reject AI altogether, but rather to ask the right questions:
- What is the expected environmental, economic, or social benefit of adopting an AI system? Are these benefits hypothetical, or have they already been rigorously demonstrated?
- Is this benefit sufficient to justify the impacts of developing, training, deploying, and continuously using the system, even though some impacts are not yet fully understood?
- Does the acceleration enabled by this AI lead to a net reduction in environmental pressures, or does it instead result in their wider proliferation?
These trade-offs require a rigorous approach based on the measurement, traceability, and transparency of environmental data. Today, it is no longer enough to make optimization promises: we need to quantify them, compare them with the actual environmental costs, and integrate them into a broader sustainability strategy.
Author: Brice Gay
Contributors: Laurin Boujon & Tristan Coignion