AI and the Environment: When the Flip Side Erodes the Coin

environment
ai
Author
Published

October 1, 2026

Presentation
Video
Extra

1 Introduction

Artificial intelligence (AI) is now presented as a potential driver of the ecological transition, capable of optimising energy systems, improving environmental monitoring, and rationalising the use of natural resources. However, this representation reflects only one side of the phenomenon: AI also relies on substantial physical infrastructure—including data centres, electrical grids, cooling systems, and mineral extraction supply chains—whose environmental footprint constitutes an issue in its own right. The report Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, published on 3 June 2026 by the United Nations University Institute for Water, Environment and Health (UNU-INWEH) in collaboration with the Government of Canada, documents this tension through quantitative data on the carbon, water, and land footprints of AI.

This is not an argument against AI, but rather a call for the responsible use of a technology whose benefits for billions of people are not disputed. This clarification invites us to adopt a measured approach throughout this article and to avoid any one-sided judgment. The analysis is divided into two parts: the first examines the documented positive contributions of AI to environmental issues (I), while the second assesses its negative environmental externalities as quantified by the UNU-INWEH report (II), before a conclusion revisits several recommendations aimed at curbing the ecological footprint of this technology.

2 AI as a tool for environmental preservation

Several areas of application make it possible to demonstrate the contribution of AI to environmental management, particularly ecosystem monitoring, precision agriculture, and the optimisation of energy infrastructure. These applications must nevertheless be presented with caution, as their actual impact is sometimes more limited than it may initially appear.

2.1 Environmental monitoring through machine learning

Machine learning applied to satellite imagery makes it possible to analyse phenomena such as deforestation and changes in biodiversity across large geographical areas, replacing partial and discontinuous monitoring with a more systematic observation of ecosystems. This capability broadens the scope of information available for public decision-making, without, however, replacing existing mechanisms for verification and governance.

2.2 Precision agriculture and the optimisation of water resources

In the agricultural sector, AI-based systems are presented as potential tools for reducing irrigation water consumption and pesticide use. These applications illustrate the potentially significant contribution of technologies capable of targeting and dynamically adjusting agricultural practices, although their widespread adoption depends heavily on the agricultural context and on local conditions of access to these tools.

2.3 Energy optimisation in digital infrastructure

The most well-documented example of energy optimisation remains the system developed by DeepMind, which achieved a 40% reduction in the energy used for cooling several of Google’s data centres. The scope of this result must nevertheless be qualified: although significant for the specific area of cooling, this reduction translates into a considerably more modest gain in the site’s overall energy consumption, as cooling represents only a fraction of the total. The UNU-INWEH report similarly calls for methodological caution by emphasising that efficiency gains, however real they may be, are generally absorbed by the growth in usage—a phenomenon known as the rebound effect or the Jevons paradox.

This final point constitutes the central analytical link of this article: it is precisely because AI is becoming more efficient and accessible that it generates growing demand, whose aggregated footprint far exceeds the savings achieved at the individual level. This dynamic leads us to examine the environmental externalities measured in the UNU-INWEH report.

3 AI as an infrastructure generating measurable environmental externalities

The UNU-INWEH report emphasises an unequivocal finding: the environmental footprint of AI is currently poorly measured because most assessments focus exclusively on the carbon emissions associated with model training. Yet every kilowatt-hour of electricity consumed by AI also carries a water footprint, linked to cooling and energy production, as well as a land footprint, associated with infrastructure and supply chains. These three footprints do not necessarily evolve in the same direction. For example, the report observes that a shift from coal to bioenergy can reduce the carbon footprint of electricity by an average of 70%, while increasing its water footprint by more than thirty times and its land footprint by a factor of one hundred. This finding leads to a measured but important conclusion: “low-carbon” electricity is not automatically “low in water consumption” or “low in land use.”

3.1 Energy consumption and carbon footprint

Global data centres consumed approximately 448 terawatt-hours (TWh) of electricity in 2025, a volume which, if considered at the scale of a country, would make them the world’s eleventh-largest electricity consumer, behind France and ahead of Saudi Arabia. The report highlights an important shift in the debate: once a model has been deployed, the inference phase—that is, its day-to-day use to respond to user queries—accounts for 80 to 90% of total energy consumption, while the initial training phase represents a more limited share.

For illustrative purposes, ChatGPT is estimated to process approximately 2.5 billion queries per day, representing around 383 gigawatt-hours (GWh) of electricity per year for this product alone.

The differences between tasks are also considerable: a standard conversational query is estimated to consume approximately 200 times more energy than a simple text classification task, an AI-generated image approximately 1,450 times more, and a short AI-generated video could consume as much energy as approximately 200,000 basic classification operations. These differences, which are often invisible to users because they are determined by default design choices, partly explain why projections for 2030 anticipate consumption reaching 945 TWh—nearly three times the combined electricity consumption of Pakistan, Bangladesh, and Nigeria, countries with a combined population of more than 650 million people.

The carbon footprint associated with this projected consumption in 2030 is estimated at 399 million tonnes, a volume that would require the planting of approximately 6.7 billion trees grown for ten years to offset, or roughly twice the estimated number of trees in the United Kingdom.

3.2 Water pressure and land footprint

The water footprint associated with the projected electricity consumption of data centres in 2030 is estimated at 9.3 trillion litres, a volume equivalent to the annual basic domestic water needs of all 1.3 billion inhabitants of Sub-Saharan Africa. This figure must be interpreted carefully: it represents a comparison of aggregated volumes at a global scale rather than direct water withdrawal from that region, but it illustrates the scale of the resources mobilised by global digital infrastructure.

The associated land footprint is estimated at more than 14,500 km², or approximately twice the area of the Jakarta metropolitan region, which is home to more than 32 million people.

The report also documents concrete local tensions. In Ireland, data centres accounted for 21% of total measured electricity consumption in 2023, exceeding the consumption of all urban households combined. This led the national grid operator to suspend new connection approvals around Dublin until 2028. Similar cases are reported in Querétaro, Mexico, where the expansion of computing infrastructure places additional pressure on water reserves during periods of prolonged drought, and in Uruguay, where a data centre project coincided with a drought that rendered tap water in Montevideo unfit for consumption in 2023.

These examples demonstrate that environmental costs are often concentrated in territories that do not directly benefit from the digital services they host.

3.3 The unequal geographical distribution of benefits and risks

The report establishes that only 32 countries host data centres specialised in AI, and that more than 90% of this global capacity is concentrated in two countries: the United States and China. Meanwhile, more than 150 countries currently have no sovereign computing infrastructure dedicated to AI.

The report describes this concentration as a digital divide with not only economic but also environmental implications: countries excluded from this infrastructure often bear the impacts of critical mineral extraction and electronic waste processing without benefiting from the associated strategic advantages.

AI infrastructure could therefore generate up to 2.5 million tonnes of electronic waste per year by 2030, a volume comparable to the annual disposal of approximately 250 Eiffel Towers.

4 Conclusion: A few recommendations

The UNU-INWEH report puts forward a set of recommendations that should not simply be accepted through intellectual laziness. According to the United Nations experts, the development of AI within planetary boundaries could be achievable, provided that technological progress is accompanied by rigorous measurement, greater transparency, and a sharing of responsibilities among all actors within the ecosystem.

This conclusion takes the form of a roadmap structured around six principles: transparency, efficiency by design, equity and environmental justice, accountability throughout the life cycle, international cooperation, and sustainable use.

We shall refrain here from assessing the relevance of these recommendations due to constraints of time and space. The issue of AI, it seems to us, is above all a democratic and human one. With regard to a technology as transformative as this, how can it be explained that peoples are not consulted to determine whether the cost-benefit balance of this phenomenon would be acceptable to them? Without disregarding the mechanisms of representation inherent in modern democracies, is it too much to ask that, on issues as consequential as AI and the environment, the direct consent of citizens should be required?

More importantly, the fact that these issues transcend the sovereignty attached to the borders of individual states implies that decisions concerning them must take into account spatial and temporal forms of otherness. This makes arguments advocating any form of purely national digital or ecological sovereignty appear both inadequate and disconnected from reality. This is all the more true because the consequences of AI—not only its environmental consequences—affect the human person in one of their most fundamental aspects: their dignity.

To be convinced of this, it is sufficient to refer to the advisory opinion of the International Court of Justice of 23 July 2025: “The implementation of the obligation to respect and ensure the right to life, and in particular the right to live with dignity, depends, among other things, on the measures taken by States parties to protect the environment.”

If cooperation among states appears to be the appropriate scale at which to address environmental challenges, and particularly those arising from the development of AI, international law appears to be the most appropriate instrument for guaranteeing respect for human dignity.

Will it succeed? Nothing is less certain, given the setbacks it has experienced in recent years, perhaps more than ever before. Its reform appears to be a necessary condition for its ability to effectively confront—undoubtedly at the cost of significant compromises—the challenges imposed by the continuation of human life on Earth, rather than the survival of the Earth itself.

Citation

BibTeX citation:
@inproceedings{madimba2026,
  author = {Madimba, Joseph},
  title = {AI and the {Environment:} {When} the {Flip} {Side} {Erodes}
    the {Coin}},
  booktitle = {Proceedings of the 1st Conference Cycles of Cytopia},
  date = {2026-10-01},
  url = {https://cytopia.fr/madimba2026},
  isbn = {TBA},
  langid = {en}
}
For attribution, please cite this work as:
Madimba, Joseph. 2026. “AI and the Environment: When the Flip Side Erodes the Coin.” Proceedings of the 1st Conference Cycles of Cytopia, accepted, October 1. https://cytopia.fr/madimba2026.