Not all artificial intelligences are equal when it comes to their ecological impact. A study reveals significant differences between models. And the smarter they are, the more CO2 they emit.
Artificial intelligences like ChatGPT, Gemini, or Claude are now part of everyday life. But behind their ease of use lies a very real energy consumption. Powering these models requires powerful servers and massive data centers, often located in the United States or Europe. Between cooling, computing, and storage, their operation is not without consequences for the environment.
A study conducted by the Hochschule München University in Germany shows that some models emit up to 50 times more CO2 than others to answer the same question. The researchers compared 14 artificial intelligences of different sizes, between 7 and 72 billion parameters. The result: the smartest models, which provide complex and detailed answers, are also the most polluting.
AI models that reason more pollute up to 50 times more than those that get straight to the point
To generate an answer, an AI transforms each word into a series of data called “tokens”. Advanced models insert thousands of “reasoning tokens” to refine their answer. This particularly concerns models like GPT-o1 and o1 mini, Claude 4.0 Opus or Gemini 2.5 Pro, present in advanced functions like Gemini Live or Deep Search.
Conversely, AIs like GPT-3.5 or Gemini Nano generate simpler answers, with much fewer calculations. As a result, the former consume much more energy. The Cogito AI, for example, achieved a good accuracy rate (nearly 85%), but generated three times more CO2 than other models of the same size.
The more accurate an AI, the more resources it consumes. The study shows that no model that generated less than 500 g of CO2 exceeded 80% correct answers. The subject matter also plays a role: answering a philosophy or algebra question requires up to six times more energy than a simple question. Researchers encourage people to prioritize concise answers and to avoid the most powerful models for basic tasks. A more "sober" use of AI remains possible, provided the tool is carefully chosen according to the need.
0 Comments