Presented as a driver of optimisation and sobriety, could artificial intelligence paradoxically worsen the environmental crisis? Behind its promises of efficiency, it hides a colossal energy and hardware footprint. The swift rise of AI over past years has been based on large-scale energy-hungry infrastructure. Training the most advanced models needs billions of computations, requiring considerable computing power. The carbon footprint of these systems is based on two key components: the power needed to run them and the physical resources required to manufacture computer components. In France, the digital sector already accounts for 10% of electrical consumption, with estimated annual growth of between 6 and 9%. Globally, digital contributes to 4% of CO2 emissions, rising constantly since 2010. “Two phases are particularly resource-hungry: training models, which can require several months of intensive computing on platforms consuming up to 30 megawatts continuously, and inference, i.e. the use of these models, which accounts for even higher energy expenditure in the case of generative AI,” explains Denis Trystram, researcher at LIG* and professor at Grenoble INP - Ensimag, UGA. “For example, training the largest models requires approximately 1026 operations, i.e. the equivalent of several months’ continuous use of the biggest platforms.”
In addition to electrical consumption, AI requires a considerable quantity of raw materials. Manufacturing electronic chips and components requires rare metals that are expensive to extract in terms of both energy and water. At the same time, data centre production and operation increase pressure on water resources in addition to their electrical consumption. Life cycle analysis of AI services is still too often incomplete, taking account of only direct effects, without considering rebound effects or systemic impacts.
Though AI can contribute to reducing certain consumption levels for specific applications - for example by optimising buildings’ heat efficiency - these gains are often cancelled out by the exponential increase in its use. Infrastructure with improved energy efficiency is not enough to offset continuous infrastructural expansion. This phenomenon, commonly referred to as the rebound effect, raises the question of the genuine ecological benefits of AI.
Certain potential areas are being explored to make AI more frugal: using data centres powered by renewable energies, developing less computing-hungry algorithms, improving the efficiency of electronic components, etc. But these solutions do not question the rampant growth of these technologies. For example, if AI is used to reduce carbon emissions, it could indirectly encourage higher consumption of other resources (such as water consumption or metal extraction), cancelling out the benefits gained.
For Denis Trystram, the real question to raise is about uses. “Do we really need everything we’re developing? This question goes beyond the hard sciences to the human sciences, such as sociology, philosophy, and economics. But these aspects are still not taken sufficiently into account in research into artificial intelligence.”
Researchers are becoming increasingly aware of the planet’s limits and the urgent need to take action, but implementing practical solutions remains complex. The digitalisation of our societies acts like a system in uncontrollable expansion, driven by the myth of unlimited growth, and absorbing other systems without thinking about the need for the uses developed.
Artificial intelligence has transformed our ways of searching for information and interacting with the world. Since the appearance of chatbots in 2022, a significant decrease in traditional internet traffic has been observed, a sign that these tools are increasingly replacing traditional searches. But this change does not guarantee a qualitative improvement of the knowledge produced. AI is based on probabilistic models whose results must be verified, and the lack of transparency of these generation mechanisms often raises questions about the reliability of the information provided. “If no strict regulation is put in place, we are heading towards saturation of resources and impoverishment of the data used to train AI. In the long-term, these models could be based only on content generated by other AI’s, creating a vicious circle where information progressively deteriorates.”
Denis Trystram uses a strong image to sum up the situation: the acceleration of our lifestyles, doped by technological progress reduces human behaviour to that of a hamster, condemned to run ever faster without every questioning the direction of travel. Is frugal AI possible? Yes, provided we get off this wheel and adopt an approach that places need before performance.
*CNRS / UGA / Inria / Grenoble INP - UGA
