a first methodology to measure the environmental cost of this technology

How can we ensure that artificial intelligence services are not too energy-intensive? The French ecosystem offers precise criteria to guide the choices of AI designers and users.

Writing texts, creating professional presentations, making images, more or less realistic videos or even turnkey songs, artificial intelligence (AI) tools have been making their way into our computers for a few months now. With simplified interfaces so that this technology – which dates back to the end of the 1950s – is now accessible with mainstream tools like Google Gemini, ChatGPT or Copilot.

To function, these AI mechanisms need access to large volumes of data, significant computing power and adequate storage capacities. Thus, a single query on ChatGPT consumes 10 times more energy than the same query on a traditional search engine. While traditional industries (automobile, aeronautics, etc.) are now working hard to modify their production methods to specifically decarbonize their activities, it would be a matter of integrating the principles of environmental requirements without delay from the design of AI services.

The need for computing for AI has increased 1 million-fold in 6 years and is increasing tenfold every year.

Sundar Pichai, CEO of Google

speech of May 14, 2024

It is necessary to create precise and quantified indicators to measure this overall ecological cost. To bring out what we callFrugal AI. However, until now, there was no internationally accessible and usable reference system for everyone. Hence the importance of the French initiative of the Ecolab of the General Commission for Sustainable Development and the French Association for Standardization (AFNOR), which published in the summer of 2024 a operational methodology to assess the environmental impact of AI.

This methodology was developed jointly by contributors from the world of academic research, businesses, associations and public administrations. This contributes to a cross-functional and non-partisan approach. So many criteria (water and energy expenditure, data storage methods, quality of datasets, reuse of already trained algorithms, etc.) that will help AI producers measure – and therefore reduce – the impact of their solutions. This will allow them to display their performance in this area, and AI buyers and users to evaluate and compare the methods of their providers.

This is a virtuous approach for all stakeholders.