AI: Friend or foe of our environmental concerns?

This article was written by Ghislain De Pierrefeu, Partner Wavestone and AI expert. The article was originally published in the daily newspaper “Le Monde”.

COP28 (COP 28 was the United Nations Climate Change Conference taking place in Dubai, United Arab Emirates, in nov-dec 2023) enthroned artificial intelligence (AI) as a tool in the fight against climate change, launching the “AI Innovation Grand Challenge, which aims to identify and support AI-powered solutions in developing countries.”

At the same time, the AI giants have stopped reporting on the energy and natural resource consumption of their data centers since the use of generative AI exploded (100 million [weekly active] users of ChatGPT). Yet we know from the few public studies available that a conversation with ChatGPT would consume the equivalent of a bottle of water, and that GPT-3.5’s training alone would have cost the carbon equivalent of 136 round-trips between Paris and New York!

Silence from the giants on their consumption

So, is artificial intelligence the unavoidable ally, or the undeniable enemy, of our environmental concerns? We often compare data with oil, and the generative AI revolution with that of the internal combustion engine. This is true enough, given the value they are already creating and promise to generate tomorrow, as evidenced by the financial valuations of AI players. Unfortunately, it’s also true in terms of negative results, like Meta’s data center project in Talavera de la Reina, Spain, which is set to use 665 million liters of water a year from a region already under water stress.

There are, however, two major differences. Firstly, we can no longer say we aren’t aware. and we have every right to be dismayed by the silence of the AI giants on their consumption. Secondly, unlike oil, there is no shortage of data on the horizon – quite the contrary. The Internet of Things, social networks and e-commerce are producing data exponentially, infinitely faster than our planet can produce oil.

AI utilizes resources for the production of chips, data storage, model training, and for the data it generates.

A real breakthrough

Most current AIs (anti-fraud, marketing, maintenance) use simple models requiring resources close to other digital uses. Natural language and image processing have already had more significant impacts, due to the complexity of understanding these human productions, often requiring the use of neural networks. However, Generative AI represents a real uptick in terms of consumption and impact, for three reasons.

First, to move towards so-called “general” AI, models need to be trained on all the data in the world. To do this, the layers and parameters of the neural network need to be multiplied, which requires titanic capacities. The second is to generate an original response for each consultation, with each word or pixel mobilizing the entire neural arsenal – which is a bit like using a specific space launcher for each satellite bolt.

Public demand

The third is the public’s appetite for these new assistants, creating a new form of usage that may well become as widespread as that of cars, search engines or smartphones. At a time when primary needs such as food, access to water, transportation and even housing are becoming luxuries for part of humanity, we have every right to wonder whether talking to an AI – however brilliant – is a sensible way to save our planet.

The environmental impact of artificial intelligence in question

We should note there are concrete benefits of AI on the environment as well, from smart energy networks to renewable energy planning, multimodal transport or waste sorting. More recently, the work of Claire Monteleoni, Computer Scientist at the University of Colorado Boulder and holder of the Choose France AI Chair, has demonstrated that the use of data from physical models combined with a neural network makes it possible to predict the trajectory of hurricanes, complex phenomena that are very difficult to predict.

Transparency

There is still a question mark over the positive contribution of generative AIs such as ChatGPT, because today we see their gimmicky and sometimes negative sides. But, looking ahead, the potential of these AIs, with their ability to propose solutions reconciling mounds of heterogeneous data, just when the physical sciences are struggling to bring together heterogeneous domains and data, is enormous.