As has long been the case with digital technologies, the environmental impact of AI is under intense scrutiny. The start of 2025 has seen a number of changes. So what exactly is the current state of play?
Here we face a real paradox, as AI, like digital technologies, can help to reduce our economy’s environmental footprint. Digital solutions are already particularly relevant in flow management (energy, water, etc.), transport, construction, industry, agriculture and local and regional authorities (waste management, lighting, etc.), as shown by most of the innovations presented at Pollutec over the years(1).
Like digital technologies, but on a much larger scale, especially when it comes to generative AI, artificial intelligence does have a greater environmental and climate impact due to its appetite for electricity, water, abiotic resources (minerals and metals), soil and seabeds (for data centres and networks) and the waste it generates.
UNCTAD’s(2) Digital Economy Report 2024 details these impacts (see box for some key figures) and urges the international community to implement global policies that promote a circular digital economy, while minimising the impacts and reducing the digital divide. The digital economy is far from being virtual. It has an underlying materiality, as highlighted by the French Agency for Ecological Transition (ADEME), the French National Centre for Scientific Research (CNRS) and the French National Institute for Research in Computer Science and Automation (INRIA), which launched the Alt IMPACT(3) programme in 2023. To date, equipment (terminals: TVs, computers, smart phones, etc.) have accounted for 50% of the environmental impact from digital technologies, with 46% from data centres and 4% from networks. This proportion is expected to change, or even reverse, with the development of generative AI and cryptocurrencies.
The impact of digital technologies on the world in numbers (Source: UNCTAD)
– The ICT sector produced between 1.5% and 3.2% of global GHG emissions in 2020
– Producing a 2kg computer requires approximately 800 kg of raw materials
– Demand for minerals such as graphite, lithium, and cobalt could surge by 500% by 2050.
– Data centres consumed 460 TWh of electricity in 2022, with consumption expected to double by 2026.
– Digital-related waste increased by 30% between 2010 and 2022, reaching 10.5 million tonnes globally, of which only 24% was collected in 2022. In developed countries, we produce on average 3.25 kg per person
– The number of semiconductor units quadrupled from 2001 to 2022 with 5G coverage expected to rise from 25% in 2021 to 85% by 2028.
– The number of Internet of Things (IoT) devices are projected to grow from 16 trillion in 2023 to 39 trillion by 2029
– E-commerce sales have jumped from USD17 trillion in 2016 to USD27 trillion in 2022 across 43 countries.
According to an ADEME-ARCEP study from January 2025, in France digital technologies represent 4.4% of the country’s carbon footprint (the equivalent of emissions from heavy goods vehicles) compared with 2.5% in 2020 (the equivalent of the waste sector) and is responsible for 11% of the country’s electricity consumption.
*ADEME and ARCEP have worked together on this matter on behalf of the French government since 2020.
How France is tackling the impact of digital technologies
In February 2020, the French Anti-Waste Law for a Circular Economy (AGEC) proved a game-changer in extending the lifespan of digital products and equipment by launching a repairability index and extending the legal guarantee of compliance for refurbished products, offering a 5-year minimum availability period on spare parts for small digital equipment, and providing information on supply times for digital equipment updates.
That same year, the country’s Digital Council (CNNum) and High Council on Climate (HCC) published a ‘Roadmap on the environment and digital technology – 50 measures for a French and European agenda on responsible digital technologies: sustainable and at the service of the ecological transition and of the sustainable development goals.’ These measures are split across three main pillars: Digital Sustainability; Digital Technologies that Contribute Toward the Ecological and Solidarity Transition; Tools and Levers for Responsible Digital Technologies.
Fast-forward to February 2021, when the French Ministry of the Ecological Transition and the State Secretary for the Digital Transition released their own ‘Digital Technologies and the Environment’ roadmap, which “ushered in a new era of public policy, centred on bringing together the environmental and digital transitions.” The roadmap includes fifteen action sheets, split across three pillars: Knowledge is Power (improving knowledge regarding the environmental impact of digital technologies); Supporting Digital Sustainability (reducing the environmental impact of digital technologies used in manufacturing, and diversifying uses); Innovate (leveraging digital technologies to support the ecological transition). Its implementation is overseen by the High Level Committee for Green Digitalisation, established in 2022.
In November 2021, a law was passed, aimed at reducing the digital sector’s environmental footprint (REEN), which “aims to bring together the digital and ecological transitions through greater stakeholder accountability within the digital sector (professionals, public sector, consumers).” This new law has five aims: to raise awareness of the environmental impact of digital technologies; reduce the amount of new digital devices; adopt eco-responsible use of digital technologies; promote more energy-efficient data centres and networks; promote a responsible digital strategy at regional level. To comply with this law, in May 2024, Arcep(4) and Arcom(5), in collaboration with ADEME, published a second version of their Digital Services Eco-design Guidelines (RGESN), which now includes 79 specific criteria.
In July 2023, an initial digital decarbonisation roadmap was created, in line with article 301 of the French Climate and Resilience Act. Resulting from discussions among several dedicated working groups (terminals, cloud and data centres, sobriety and uses, networks, the role of digital technologies in the decarbonisation of other sectors), it identified challenges to overcome and actions to be taken in order to achieve decarbonisation targets by 2030, then 2050.
A ‘Digital Eco-responsibility’ acceleration strategy was adopted alongside this as part of the France 2030 initiative to strengthen eco-responsibility within the digital sector while also offering more sustainable yet competitive solutions. The strategy has four pillars: supporting methodological developments to strengthen knowledge of the environmental footprint of digital technologies , as well as research into the eco-design and energy-efficiency of digital solutions; promoting innovation for a circular economy in the digital sector , in the process making France a leader in eco-design, energy-efficiency, and lifespan extension of digital products; creating (initial and continuous) training opportunities relating to digital eco-design and energy-efficiency; acclimatising the various stakeholders to this digital transformation and supporting them throughout the process.
However, although as covered above, France benefits from unique legislation, ADEME considers that significant efforts are still required to understand and disseminate the impacts of digital technologies. It therefore recommends extending the useful life of our equipment (eco-design, repair, refurbishment and recycling), encouraging mindful use, controlling the rebound effect(6) of network technologies and innovating to anticipate future impacts, in particular AI and virtual worlds.
Defining AI
Before scrutinising the environmental footprint of artificial intelligence, it’s worth remembering what it really means.
The French Data Protection Authority (CNIL) defines AI as being a “logical, automated process generally based on an algorithm and capable of carrying out well-defined tasks.” The CNIL goes on to say that for the European Parliament, artificial intelligence constitutes “a machine’s ability to display human-like capabilities such as reasoning, learning, planning, and creativity”. It also adds that: “The European Commission considers that AI encompasses machine learning approaches, logic and knowledge-based approaches and statistical approaches, Bayesian estimation and search and optimisation methods. ”
Generative AI, on the other hand, refers to any artificial intelligence that generates content (including text, audio, video) from data. And when it comes to AI, data is key! Whether its models are being trained or fully operational, AI needs to process a considerable amount of data in order to function. The International Energy Agency (IEA) estimates that electricity consumption by data centres will be over 75% higher in 2026 than in 2024 (i.e. more than 800 TWh vs 460 TWh), resulting in an increase in GHG emissions. Additionally, the current boom in generative AI will lead to the release of new better-adapted equipment (computers, smartphones, tablets etc.)
Factoring in AI’s environmental footprint
Issues relating to the environmental footprint of artificial intelligence have recently started to be addressed on a large scale. As an example, in June 2024 Ecolab(7), the Association Française de Normalisation (AFNOR) and ADEME published their general reference framework for frugal AI(8). This introductory guide to measuring and reducing the environmental impact of AI systems sets out definitions, methods and best practices for the promotion of a frugal approach to the creation and use of AI services. The guide lends itself to AI developers as well as public and private buyers wishing to adopt a sustainable and responsible approach.
In December 2024 ADEME and ARCEP launched the Observatory of digital technology’s environmental footprint, in accordance with article 4 of France’s REEN law. The objective of this Observatory, which also encompasses AI, is to “establish a benchmark platform for reliable and corroborated data on the environmental impacts of digital technologies. ”
The AI & Society Institute, established by the École normale supérieure (ENS-PSL) and the ENS Foundation, launched another Observatory in early February 2025 in collaboration with Capgemini. The Observatory focuses primarily on analysing AI’s environmental impacts at all stages of its life cycle (training, fine-tuning, inference and end-of-life), as well as their mitigation. The objective is to develop “a robust and shared methodology to promote sustainable AI usage practices. ”
On a global scale, the International Energy Agency (IEA) will be instigating an Observatory on Energy, AI and Data Centres. Due to start work in April, the Observatory intends to gather data on AI’s hunger for electric power, in addition to tracking cutting-edge AI applications across the energy sector and anticipating the energy needs of data centres and AI models.
Moving towards the environmental sustainability of AI worldwide
At February’s AI Action Summit hosted in Paris, 61 countries, along with the European Union and the African Union Commission, signed a “Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet.” The signatories, which did not include the US or the UK, consider it crucial to ensure AI is sustainable for both people and the planet.
Alongside this, a Coalition for Sustainable Artificial Intelligence has been launched by the French Ministry for the Ecological Transition to give AI sustainability the same prominence as security and ethics in the global debate on AI. Supported by ten countries and six international organisations, this Coalition includes many leading figures from across the AI ecosystem (large companies, research centres, associations and NGOs, investors, public institutions, SMEs and start-ups), all committed to setting AI on a more responsible path. Watch this space.
1) For example: solutions for optimising processes, monitoring (sensors, etc.), preventative maintenance, and so on.
2) UNCTAD: UN Trade and Development – https://unctad.org/publications
3) Created by the French Agency for Ecological Transition (ADEME), the French National Centre for Scientific Research (CNRS) and the French National Institute for Research in Computer Science and Automation (INRIA), the programme Alt IMPACT – se mobiliser pour la sobriété numérique [Working towards digital sobriety] – includes a website for individuals, businesses, and local authorities – https://altimpact.fr/en/
4) ARCEP: French Electronic Communications, Postal and Print Media Distribution Regulatory Authority (previously ART, founded in 1997, and renamed ARCEP in 2005)
5) ARCOM: Regulatory Authority for Audiovisual and Digital Communication, formed from the January 2022 merger of the High Audiovisual Council (CSA) and the High Authority for the Distribution of works & Protection of Rights on the Internet (HADOPI)
6) Rebound effect: When certain environmental gains achieved by more efficient resource management or technical developments are significantly reduced or cancelled out by an increase or a shift in usage (France Terme – a French terminology base). The rebound effect can be direct (shifting efficiency gains from one activity within that activity) or indirect (shifting gains to another activity).
7) Ecolab – Greentech Innovation – is the innovation lab for the ecological transition within the French General Commission for Sustainable Development (CGDD)
8) AFNOR SPEC 2314 – An AFNOR SPEC to measure and reduce the environmental impact of Al
Read more:
Report by the French Economic, Social and Environmental Council (ESEC): “Impacts de l’IA : risques et opportunités pour l’environnement” (in French): Fabienne Tatot and Gilles Vermot Desroches, 60 pages, ref. 2024-014, September 2024. The report addresses AI for the environment (“AI for green”) and frugal AI (“Green AI”). It sets out six avenues for developing frugal AI for the environment.
Report by the ADEME: “Numérique & environnement : entre opportunités et nécessaire sobriété” (in French): Erwan Fangeat and Mathieu Wellhoff, 12 pages, ref. 012717, January 2025. The report’s proposed recommendations mainly concern limiting numbers of devices, reducing usage, measurements, controlling rebound effects, and innovation.
