“One of the structural characteristics of AI is that it can take into account a multitude of factors, which adapts very well to environmental issues, which are essentially multifactorial,” says Gilles Babinet in his book GREEN IA.
Just like digital technologies, AI is very often put forward as a new means, a solution to decarbonize society and the current economy.
Many tools are already using AI to put it at the service of the ecological transition. Often, its digital technologies are described as mitigating climate change on issues such as improving energy management by increasing energy efficiency or promoting the adoption of low-carbon technologies.
But these energy gains must be systematically observed in the light of a cost-benefit analysis that includes the entire life cycle of the digital technology used.
Consideration should also be given to the increased demand for goods and services due to the use of digital devices, which can greatly reduce the gains resulting from certain technologies. Indeed, AI is followed very closely by its rebound effect. The rebound effect applies particularly well to digital technology because it is governed by rare exponential laws. For example: microprocessors that double in power every two years, the development of very demanding uses in terms of calculation, the importance of data flows, the frequent incentives to change smartphones, which is the largest source of CO2 emissions in the digital industry.
And followed even more closely by the notion of the effect of accelerated obsolescence, which must be avoided as much as possible. It is a question of replacing an object that functions solely by the desire to have objects that integrate AI. These effects are in total contradiction with the logic of frugality and sobriety.
The ADEME ARCEP report exposes the fact that the impact of digital technology is likely to triple by 2050 to reach 6.7 percent of France’s carbon footprint, but this report does not take into account the explosion in the use of AI. It is reasonable to think that the use of AI is likely to increase the environmental footprint of digital technology, which is booming.
For example, ChatGPT requires 4 to 5 more computations than a traditional search engine to generate its responses. The simple training of ChatGPT released 552 tons of C02 (the equivalent of 205 round trips from Paris to New York by plane). According to ADEME, ChatGPT’s inference generated a daily footprint of 23.04kg CO2. In sum, a ChatGPT query generates a total of about 1.54g of cO2.
Indeed, its specific properties, represented by its computing power through large algorithmic models, have a colossal environmental impact. In terms of energy, the data centers where the calculations are carried out, which consume a lot of energy since a large amount of water is needed for their cooling system. But all the more so by the use of GPUs (processors present in AI that are particularly energy-hungry and are composed of rare earths).
In view of the implications of deploying an AI system, it should be weighed directly against environmental, legal, ethical and business skills issues.
The fundamental step in the implementation of such a system within one’s organization is to question the need for and relevance of an AI system. It is necessary to systematically start from the need, identify the AI’s potentials (classification, regression, optimization, generation) and then study whether one of the potentials is related to the previously defined needs.
Secondly, would a traditional digital tool be enough? Knowing that in some use cases, a smaller algorithm can do the trick very well. For example, to make predictions, a linear regression model through statistical modeling can be more than enough.
