An average user of generative artificial intelligence (AI) rarely considers the environmental impacts of a digital interaction with a Chatbot, due to the seamless experience that AI interfaces provide. In reality, to generate a 100-word email, ChatGPT “drinks” about 500 millilitres of water, roughly the amount contained in a 16-ounce bottle. Additionally, a single prompt consumes nearly ten times more energy than a standard engine query such as Google.
Source: ChatGPT.
In recent years, the rise of AI has generated a lot of excitement due to its potential to enhance efficiency and decision-making across a wide range of sectors. However, as the models become more advanced and complex, they require larger datasets and greater computational power, placing a heavier strain on our planet’s natural resources. With the increasing use of tools such as ChatGPT, it is essential to consider their energy demands and the broader implications for environmental sustainability. While the full extent of AI’s environmental footprint remains difficult to evaluate, this article highlights some risks and discusses how policymakers and legal frameworks can
help mitigate them.
Source: Canva.
The life cycles of generative AI
A comprehensive assessment of AI’s environmental impact requires examining its full life cycle, covering both the software and hardware phases. On the one hand, the software cycle involves data collection, model training, deployment, and ongoing maintenance of the systems. The hardware life cycle, on the other hand, involves the production of physical components such as chips, including graphical processing units (GPUs), which are essential for training AI models and powering data centers.
The Surprising Electricity Demand of AI Tools
As stated by the Chatbot itself in the introduction, AI systems consume a staggering amount of electricity, leading to an important increase in the emissions of planet-warming greenhouse gases. What sets generative AI apart from other systems is its significantly higher energy demands. During the software cycle, training large generative AI models such as OpenAI’s GPT-4, which contains billions of parameters, requires an immense amount of computational power. This results in high electricity consumption, which in turn increases carbon emissions and puts pressure on the power grid. The International Energy Agency’s Energy and AI report predicts that by 2030, the demand from AI-optimised data centres will have quadrupled.
Dependence on Critical Minerals and Unsustainable Mining Practices
In addition, dependence on critical minerals has become a major concern in this digital age. Large-scale AI deployments, usually hosted in cloud-operated data centres, rely heavily on them during the hardware production cycle, particularly in the fabrication of microchips. The extraction of rare earth elements often involves environmentally destructive practices such as water contamination, soil degradation, habitat destruction and further ecological harm. Moreover, growing demand for semiconductors is straining already fragile supply chains which are threatened by China’s quasi-monopoly over the rare earth minerals market.
Electronic Waste from AI Infrastructure
Moreover, data centers heavily rely on non-renewable energy sources to power devices, cool systems, and maintain the overall infrastructure. The energy used to fuel AI systems at the hardware phase often comes from the burning of fossel fuels which produce planet-warming greenhouse gases and an astonishing amount of electronic waste (e-waste). A study led in 2019 estimated that training a large language model (LLM) produces around 300,000 kilograms of carbon dioxide which is the same amount of energy used for 125 round-trips flights between New York and Beijing. In addition, due to the exponential advancement of technology, the hardware used to develop AI systems often has a short life span, leading to frequent disposal. When the materials are discarded, they release substances such as lead and mercury which pose serious risks to both human health and the environment.
The Hidden Water Cost of AI
One of the most surprising and often overlooked environmental impacts of AI is its water consumption, particularly during the cooling process. What many people don’t realize is that AI models continue to use water even after training is completed. In fact, about two liters of chilled water are needed for every kilowatt hour (i.e. about the amount of energy used by a typical household appliance in an hour) of energy consumed to cool down the systems and prevent them from overheating. A study conducted by Cornell University demonstrated that training ChatGPT in U.S. data centers causes the evaporation of 700,000 liters of clean and fresh water. This is particularly concerning given that a quarter of the global population already lacks access to clean water and that fresh water is becoming a scarcer resource. Additionally, in warmer and drier regions where water management is already challenging, the additional demand of water for AI models creates tension between water needs of humans and data centers.
Can Legal Frameworks Rise to the Challenge?
Despite the many concerns highlighted in this article, effective legal frameworks and robust policies can help limit AI’s environmental footprint in the long run.
1. Addressing the environmental footprint through legislation:
- Many states have already demonstrated their awareness of the issues listed above by adopting legislation aimed at mitigating the environmental footprint of AI.
- For example, in the United States, congress adopted the Federal Artificial Intelligence Environmental Impact Act on 1 February 2024, which addresses the environmental challenges posed by emerging technologies. This bill requires the administrator of the Environmental Protection Agency to regularly carry out studies on the impacts of AI.
- The European Union responds to these concerns directly through the AI Act. In fact, article 37 provides that “the European Union must consider the protection of the environment in its policymaking”.
- Meanwhile, in Canada, Bill C-27 was supposed to address similar concerns, but it did not come into force following Justin Trudeau’s resignation.
2. Promoting an effective management of e-waste through policy:
- The United Nations Environment Programme (UNEP) has made several recommendations emphasizing the need for governments and policy makers to intervene promptly to minimize the long-term consequences of AI on the environment.
- According to UNEP, governments should implement regulations requiring companies to disclose the environmental impact of AI products and services, including metrics such as water and energy consumption needed to operate these systems.
- It has also been suggested that existing policies such as extended producer responsibility (ERP) can be used to hold manufacturers accountable for the environmental impact of their products throughout the different cycles of AI. Such an approach could ensure that they adopt more sustainable practices in the long run.
In conclusion, despite evident concerns surrounding AI’s impact on our planet’s health and the limited effects of current legislative efforts, continuous intervention by policy makers could make an important impact in the long run.
Ce contenu a été mis à jour le 29 octobre 2025 à 15 h 39 min.