The prisoner’s dilemma is a classic of game theory that shows how hard it is to cooperate, even when it’s in everyone’s interest. Two people are separated, they can’t communicate, and they’re offered a deal. If both stay silent, both get a light sentence, six months in prison for example, because there’s no evidence. If one gives up the other and the other stays silent, the one who talked walks free, and the other gets a heavy sentence, ten years for example. And if both give each other up, they betray each other, five years each for example.
What’s interesting is that for each prisoner, the temptation to betray is strong: if the other stays silent, betraying gets you released right away, and if the other betrays and you kept quiet, you get a very heavy sentence. So betrayal is the dominant strategy, even though if nobody betrays, both come out winners. That’s what’s called a Nash equilibrium: each one betrays, and neither can improve their lot by changing their decision alone, even though both would be better off staying silent. To get out of it, both have to change at the same time.
LLMs vs eco-design
This deadlock is the one where, with AI, the players who care about environmental issues and who work in a highly competitive sector built on high technology find themselves today. Because obviously, and it’s perfectly undeniable, AI today has a very bad environmental impact. It uses a huge amount of electricity, it uses water (that’s up for debate: closed loop or not, rise in temperature). In any case, even if it’s the same system as in nuclear power, it very likely doesn’t have a perfectly neutral impact. And water is becoming a strategic resource, the summer of 2026 we just went through proved it. Each request weighs almost nothing, sure: around one gram of CO2 and a few dozen milliliters of water for a 400-token response, according to Mistral’s study (opens in new window). But the total is anything but negligible: data centers consume 485 TWh worldwide, and could double by 2030 to reach the equivalent of Japan, according to ADEME (opens in new window).
Alongside that, AI makes amazing things possible: making teams more autonomous, thinking about things differently, getting a taste of new skills to open up new perspectives we would never have allowed ourselves until now. Working prototypes through vibe coding, motion design or simple graphic design for those whose job it isn’t, work on SEO, 3D modeling… “Sky is the limit” for anyone who wants to experiment and isn’t too fussy about the perfect quality of the details of what’s put in their hands. We’re not going to go over everything it makes possible, there are enough demos all over the place to see the potential. What’s certain is that today, whatever the numbers, it’s unthinkable that seasoned users of these technologies aren’t building themselves added value, a competitive advantage, and it’s not unthinkable that those who don’t are putting themselves in an uncomfortable position in terms of competitiveness, unless they choose to take the complete opposite stance of fully showcasing human know-how. Not everyone will necessarily be able to do that.
High-performing or mindful
So for the players who need to perform and who are asking themselves questions on the environmental front (see the article on EcoIndex and the cobra effect), we’re in exactly this situation. The whole sector, agencies, IT services firms, freelancers, would need to ease off, but it only takes one not easing off for nobody to do it. We’re no exception, our convictions are taking a bit of a beating right now, we’re actively working on it, notably by testing local models!
Sources
- ADEME, opinion « L’intelligence artificielle générative, des impacts environnementaux importants », 2026. Data center consumption: 485 TWh worldwide in 2025, possible doubling by 2030, multiplied by 3.7 in France by 2035. https://www.ademe.fr/presse/communique-national/ia-generative-comment-quantifier-les-impacts/ (opens in new window)
- Mistral AI, “Our contribution to a global environmental standard for AI”, July 22, 2025, life cycle assessment of Mistral Large 2 carried out with Carbone 4 and ADEME, reviewed by Resilio and Hubblo. A 400-token response on Le Chat: 1.14 gCO2e and 45 mL of water. The company’s study of its own model, presented by it as a first approximation. https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai (opens in new window)
The illustrations in this article were generated by AI (see our AI use statement).
