Position Paper: five challenges for more environmentally-friendly artificial intelligence

The rise of artificial intelligence is transforming our societies at great speed. Capable of improving our understanding of environmental phenomena, optimising energy consumption and promoting ecological transition, it is emerging as a useful tool in the face of climate change. However, its development and use are accompanied by a worrying ecological footprint, due in part to massive energy consumption, intensive use of natural resources and a low rate of component recycling.

Faced with these challenges, the Coalition for Ecologically Sustainable Artificial Intelligence, announced at the AI Action Summit in Paris, aims to bring together companies, researchers and institutions to promote more sustainable AI. Bringing together 91 partners, including several major technology companies and international organisations, this initiative has resulted in the publication of a Position Paper coordinated by Inria and the French Ministry for Ecological Transition, which identifies five major challenges for reducing the environmental impact of AI and encouraging more responsible practices.

These challenges, based on recent findings and developments, have two objectives: to limit the environmental footprint of AI, but also to develop sustainable AI tools capable of enabling scientific advances and innovations. These advances go well beyond the optimisation of existing systems, and their ambition is to profoundly transform current practices.

1 - Environmentally efficient technologies

The first challenge identified by the authors is to develop more efficient AI technologies that respect the environment, by optimising hardware, algorithms and data.

Energy optimisation is essential, but other impacts, such as water consumption, also need to be considered. Advances include more efficient architectures (digital accelerators, embedded AI), optimised algorithms and new cooling techniques. Finally, bio-inspired approaches, such as neuromorphic architectures, offer promising solutions for more sustainable AI.

2 - Towards specialised, agile models trained on reliable data sets

Training AI models is extremely energy-intensive. Digital technology accounts for up to 12% of the world's electricity consumption, and this figure could rise sharply in the coming years.

The second challenge identified is therefore to develop more specialised AI models that are small and efficient, rather than relying solely on large generalist models. These targeted models make it possible to improve efficiency while reducing environmental impact.

3 - New methods and better data to assess the environmental footprint of AI

The third challenge is to better assess the environmental footprint of AI by developing precise methods and indicators.

This means analysing the entire lifecycle of AI systems, from the manufacture of the equipment to its energy consumption and emissions. The sharing of data by companies and the standardisation of assessments are therefore essential to guarantee greater transparency. In addition, open source models can encourage the sharing of resources and avoid the duplication of training models for similar uses, thereby ensuring more efficient consumption of energy and resources.

4 - Applying the principles of the circular economy to AI hardware

Currently, the manufacture and rapid renewal of equipment generate significant consumption of natural resources and little recycled electronic waste.

The aim of the fourth challenge is therefore to promote sustainability by optimising the design, use and recycling of components. This means giving preference to recyclable materials, encouraging repair and reuse, and developing responsible supply chains to limit the ecological footprint of equipment.

5 - Changing the image of AI to promote the development of frugal AI tools and their rational use

Finally, beyond infrastructures and algorithms, AI performance is currently measured mainly by quantitative criteria, which are often general, favouring large-scale systems over specific, targeted solutions.

The fifth challenge therefore calls for a change in the way we measure and evaluate performance, in order to give greater value to research into frugal, economical and even low-tech AI systems. It is also important to educate professionals and the general public about the environmental challenges of AI, and to develop incentive policies to encourage less energy-intensive AI.