Artificial intelligence has invaded our daily lives and changed the way we consume, but its dazzling development hides a far less exciting reality: its environmental footprint is growing and generating considerable digital pollution. Behind the prowess of algorithms lies massive energy consumption, fueled by Big Data and training processes that consume ever more electricity and infrastructure. This makes it all the more urgent to assess its ecological impact.
To measure the environmental footprint of Artificial Intelligence, it is essential to take into account different components of the value chain: energy, carbon footprint, water resources, as well as ores, minerals and infrastructure.
We’re heading for an explosion in energy consumption, an increase in greenhouse gases, the use of large quantities of water to avoid overheating and the fate of electronic components in our hands.
There’s nothing like numbers to make you realize this…
A modern data center consumes the energy equivalent of 100,000 homes. Their emissions are expected to triple by 2030, reaching 1.4% of global combustion emissions, a footprint comparable to that of commercial aviation. Training ChatGPT-3 would emit 240 tonnes of CO2 per year, equivalent to 136 round-trips between Paris and New York.
According to a report by the Economic, Social and Environmental Council, “developing and training ChatGPT-3 required 700,000 liters of water”. For its part, Microsoft has increased its worldwide water consumption by 34% in 1 year. At this rate, the OECD estimates that “AI could consume 4.4 to 6.6 billion cubic meters of water by 2027, half of what the UK withdraws every year”.
On the other hand, to ask him “25 prompts would cost half a liter of water, and asking for help writing a 100-word e-mail would be equivalent to wasting 519 milliliters of water, a little more than a small bottle”. With 200 million users, you can imagine the scale of the problem.
According to the International Energy Agency, “a query on ChatGPT consumes as much energy as 10 Google searches”. What’s more, generating 100 images via AI is equivalent to recharging 950 smartphones or driving a gasoline-powered car 6.6 km.
Another important point linked to Artificial Intelligence is that of e-waste management (electronic components, data centers…). One study estimates that generative AI generated 2,600 tons of e-waste in 2023. If nothing is done, this figure could rise to 2.5 million tonnes by 2030, equivalent to 13.3 billion discarded smartphones.
Solutions:
- Frugal AI, which reduces the environmental impact of AI systems by taking into account the equipment used, energy consumption, greenhouse gas emissions and the impacts linked to the manufacture and recycling of components.
- Free-cooling, which lowers temperatures by using outside air, without requiring the use of water.
- Experimentation with underwater data centers.
- Choose the right country to train your AI model: the environmental impact varies from country to country – training in France emits only 93,000 kg of CO2, whereas in India emissions would be 10 times higher.
- Google and Microsoft explore the use of modular nuclear reactors
Sources :
- IA : générer une image avec ChatGPT consomme-t-il vraiment un litre d’eau ?
- Bilan carbone de l’IA : à la fois colossal et sous-évalué – Colibris Blog
- Derrière l’essor de l’IA, une bombe énergétique que personne ne veut désamorcer
- BFMTV.com. Pourquoi l’IA générative a-t-elle besoin de tant d’énergie ?
- L’impact écologique de l’intelligence artificielle : quelle est son empreinte carbone ?
Master 2 ｜ Droit de l’économie numérique
