CSR: When AI helps with the ecological transition

More topical than ever, the need for an ecological transition is leading to many initiatives to "green" the processing and storage of data, both of which are historically energy-intensive activities. Sylvain Bouveret, head of the Information Systems Engineering programme at Ensimag, explains how Artificial intelligence (AI) can help with the transition.

Digital tech: impacts that are difficult to assess

According to research, between 2% and 4% of global greenhouse gas emissions are due to the world’s digital ecosystem – a volume twice that created by air transport. However, for Sylvain Bouveret, the comparison is of only limited value. "Unlike the aviation sector, where the impact comes mainly from the use of aircraft, in IT, the individual pieces of equipment consume relatively little energy; but it’s a ubiquitous sector and its impacts are everywhere. User terminals, data centres, telecommunications networks, connected objects... the sheer variety and number of these devices make the sector’s impact very difficult to estimate.”

What’s more, the calculations also underestimate the indirect and rebound effects of digital tech, according to the expert. "Some of the digital tools that we design are intended to accelerate and optimise business flows, but we mustn’t overlook the environmental cost that all this creates. It’s the case, for example, with logistics flows, which are increasingly being optimised by digital technology to connect producers and consumers from one end of the planet to the other, and to deliver goods in less than two days. While all this has a major impact, it’s difficult to quantify the share that can be attributed to digital technology.”

What about AI in all this?

When it comes to the environment, Sylvain Bouveret points out that AI’s impact is mainly in the model training phase, which is resource intensive even when carried out offline. He also underlines the need to include the manufacturing of IT hardware in impact calculations. "We often focus on energy consumption linked to usage of the equipment, but the impact is just as significant at other stages of its life cycle, such as production or end of life.” By way of example, he cites a study by GRICAD, a high-intensity computing and data facility in Grenoble, France, which used different criteria to measure the carbon footprint of one hour of computing. It found that 40% of the carbon footprint of server-side computing was produced by manufacturing the hardware, hence the need to focus on the entire life cycle of IT equipment.

A mission to make AI greener

Today, many initiatives are emerging at corporate and government levels to develop a greener version of AI. In line with the European Green Deal, which aims to make the EU carbon neutral by 2050, many cloud computing companies, for example, have set carbon neutral targets for their data centres by 2030. These facilities, which are more efficient in terms of power usage effectiveness (PUE) and consume less energy, are now referred to as ‘green data centres.’ However, as Sylvain Bouveret points out: "We mustn’t forget that carbon neutrality is a very broad vision and that just being carbon-centric about this won’t solve the environmental crisis that we now face. As for AI specifically, he adds: "More and more work is being devoted to low-energy approaches and trying to find learning methods that use resources more efficiently, courtesy of dedicated architectures.” Ecoinfo, a service group within France’s National Centre for Scientific Research (CNRS) of which Sylvain Bouveret is a member, has made several recommendations for both analysing and preventing the environmental impacts of AI projects. Ecoinfo’s advice is to:

- Carry out a systematic evaluation of impacts, using proven methods such as life cycle analysis, which includes other indicators than CO2 emissions.

- Try to assess as accurately as possible the indirect and rebound effects of AI projects (e.g. an increased use of certain services).

- Introduce countermeasures to avoid these indirect and rebound effects.

Though difficult to implement, these recommendations highlight the need for a broader approach to analysis, one that includes the different impact criteria, the various stages of the equipment life cycle and the sheer complexity of these systems of terminals, data centres and networks.

AI, an ally of CSR strategies?

Although AI has a significant environmental cost, it can also help with the ecological transition and act as a solid ally of CSR strategies. Applications of the Internet of Things (IoT), for example, have been developed to measure and optimise resource consumption and management in a very efficient way.

However, Sylvain Bouveret warns that the rebound and indirect effects of AI projects should not be overlooked. "It’s essential to combine these approaches with detailed cost-benefit analyses, so that the AI solution is at the right scale, meets a specific need and isn’t a pointless waste of energy.”

For Sylvain Bouveret, one of the extremely positive aspects of AI is its undeniable ability to increase our knowledge and understanding of climate risk. But he concludes: "At a time of widespread scarcity of resources, it’s more important than ever to move to a low-energy environment in digital technology, as everywhere else."