Although AI has been around for years, it has been in the news a lot since the appearance of ChatGPt. However, the multiplication of its use cases must be framed with regard to its environmental impact, which is growing and yet less animated by debates.

 According to the Green IT report, in 2019, 34 billion devices were present worldwide for 4.1 billion users, or 8 devices per user.

In this respect, traditional digital technology alone already accounts for 2.5% of France’s carbon footprint, according to the Ademe Arceps report (1) and 4% of the world’s carbon footprint (higher than civil aviation emissions). Although for the moment modest compared to other sectors, it is envisaged that it will jump by 60% by 2040 if nothing is done to reduce it. That is to say 6.7% of France’s carbon footprint, which is in total contradiction with the respect of climate policy objectives in terms of carbon neutrality by 2050.

This carbon footprint is made up of:

  • To 70% of terminals (especially screens and televisions)
  • From 4 to 20% by data centers.
  • From 4 to 13% by networks

The environmental impact of AI is not limited to an approach to CO2 emissions, but to date it is the only existing calculation methodology.

 These technologies give off many other negative externalities. The explosion of these uses leads to an increase in the demand for abiotic resources

The use cases of digital devices and digital services are exponential, so the use cases of rare metals that are essential to them are directly correlated. This has a direct impact on their depletion (Cobalt, lithium, neodymium, indium, nickel, manganese, etc.). The drilling activities it generates also emit a lot of greenhouse gases.  The disposal of these metals and ores at the end of their life also emits its share of pollution, as the recycling channels for these metals are currently underdeveloped. The miniaturization of components makes it very difficult to recycle them.

 Two stages in the manufacture of digital equipment are particularly emitting: the extraction of raw materials and their transformation into electronic components. (It is also worth questioning the supply of these materials which, although for the moment sustainable, can generate high risks and tensions in the countries from which we source).

 For your information, the manufacture of digital equipment represents (in 2019)

  • 30% of the overall energy balance,
  • 74% of water consumption,
  • 39% of GHG emissions
  • 76% of the contribution to the depletion of abiotic resources.

 Secondly, the calculations, allowing the training and inference of AI, run on data centers, which creates a phenomenon of overheating and cooling procedures that require a significant consumption of water resources. Next to that, the demand for electricity is just as important. It should be noted that its emissions vary depending on the location of the data centers, which depends on the energy mix of the host country. Indeed, this “carbon-based” approach can be undermined in countries such as France, where 92% of electricity is carbon-free, says Gilles Babinet in his book Green IA. It is important to keep in mind that on average, data centers and large-scale calculations release 100 Megatons of CO2/year. According to the consulting firm Carbone 4, between 2013 and 2017, their energy consumption increased by 50% to almost 10% of global electricity consumption.

We have very little information about these data centers, which prevents us from carrying out full environmental assessments, as the data centers are mostly located in non-European countries. And for reasons of trade secrets, very little data is reported. (This also raises questions about digital sovereignty). To date, no evaluation methodology has been set to measure the environmental impact of these AI systems, so it is necessary to seize them quickly.

Another angle of the environmental impact of AI must be taken into account. This is the effect of accelerated obsolescence, a phenomenon of replacing equipment that works perfectly well in favor of equipment integrating AI.

A propos de Line Hinderer