The ecological transition is now a priority issue for all tech players. In this context, AI, whose use is experiencing exponential growth, raises questions about its environmental impact. During the Green Tech Forum 2024, a round table brought together experts to take stock and share best practices in responsible and frugal AI.
Participants particularly highlighted the essential role of public authorities in regulating digital technology and the need for a systemic approach involving all stakeholders. They also highlighted initiatives such as developing eco-responsible data models and infrastructures. This article reviews the main conclusions of this round table, which is crucial for the future of artificial intelligence.
The environmental footprint of AI: innovative modelling tools
Faced with the rapid evolution of artificial intelligence technologies, modelling the environmental impact of AI represents a significant challenge. At Publicis Sapiens, Vincent Villet has developed an open-source model for assessing digital systems’ environmental footprint before deployment.
This solution is based on public data from ADEME in collaboration with Boavizta. The model allows an in-depth analysis of digital systems’ functioning, integrating the study of usage curves and user journeys. It also assesses server requirements and accurately estimates the carbon impact of the services developed.
The impact varies depending on the type of AI.
The modelling highlighted significant disparities depending on the use of AI. The analysis reveals that generative AI applied to images and videos has an exceptionally high environmental cost.
In contrast, text-only applications show a much more moderate impact. One encouraging point is that systems’ energy efficiency is improving rapidly, offering higher performance today for the same consumption as a year ago.
The Frugal AI Framework: A Government Initiative
The General Commission for Sustainable Development (CGDD), through its Ecolab DataIA Innovation office, has developed a framework for frugal AI with AFNOR. This reference framework, accessible free of charge, is structured around three essential cyclical phases.
The upstream phase is a critical step in which the team assesses the relevance of using AI. This initial reflection involves an in-depth analysis of possible alternatives and guides the choice of the most appropriate model.
During the development phase, the focus is on data optimisation. Teams rigorously select relevant data and maintain continuous performance analysis to ensure system efficiency.
The post-project phase allows you to capitalise on the lessons learned. It includes establishing governance dedicated to frugality and encouraging the continuous improvement of practices. This last step is crucial for the development of future projects.
Responsible AI: Concrete applications in the territories
The France 2030 program illustrates the government’s commitment to financing frugal AI projects in the territories. These initiatives aim, in particular, at the decarbonisation of buildings, tackle the challenges of preserving biodiversity or contribute to reducing water leaks.
These projects, led by emerging companies in close collaboration with local authorities, demonstrate that technological innovation and environmental responsibility can be reconciled. They create a virtuous synergy between local economic development and sustainability objectives.
The need for dual expertise
Also, note the crucial importance of combining AI expertise and skills in sustainable IT. Traditionally focused on financial optimisation, data scientists and AI engineers now systematically integrate environmental considerations into their approach.
This evolution of professional practices illustrates the sector’s necessary transformation, where technical excellence must be accompanied by increased ecological awareness. This dual expertise is becoming an essential standard for the responsible development of AI.
Limit the volume of trade.
The environmental impact of generative artificial intelligence is measured in particular through the volume of token use. During interactions with AI models, the volume of data exchanged directly influences financial costs and resource consumption. For example, a conversation of 30 to 50 exchanges with a generative AI corresponds to a water consumption of approximately 500 ml.
Responsible approach to AI development
Experts point out that creating new AI models to meet customer needs is generally unnecessary. 95% of use cases can be covered by existing models, such as GPT, which have already been trained on massive volumes of data. To optimise the responses, several solutions are possible:
- Adding specific datasets to refine responses
- The use of RAG (Retrieval-Augmented Generation) to exploit the company’s documentary databases
- Optimizing prompts to reduce the number of exchanges required
Infrastructure and data centers
The use of AI relies on substantial cloud infrastructure, represented by approximately 8,000 data centers worldwide. According to projections by the International Energy Agency, their energy consumption in 2026 will be equivalent to that of Japan.
Innovation in responsible data centers
Innovative solutions are emerging, as illustrated by the example of Infomaniak in Switzerland:
- Data centers without air conditioning
- Underground installation to minimise landscaping impact
- Heat recovery system:
- Conversion of electrical energy into heat
- Use of heat pumps connected to the district heating network
- Cooling of servers by the generated cold flow (28°C)
- Heating 6,000 households per year
- However, the environmental impact remains significant:
- A server represents 1.7 to 2.5 tonnes of CO2 equivalent
- Each GPU generates 112 kg CO2 equivalent
Role of public policies
Faced with digital technology’s environmental challenges, public policies appear to be an essential lever for regulation. The State takes on a dual role as prescriber and trusted third party, enabling it to drive and support practice transformation. Its action is mainly part of the European regulatory framework, considered the relevant level for the common market.
At the national level, the State also acts as a facilitator, particularly by coordinating initiatives such as the one carried out with AFNOR, where it brings together the stakeholders involved to define the standards for responsible digital technology collectively.
Towards a reasoned vision of AI
The roundtable participants converged on a pragmatic vision of artificial intelligence’s role in the ecological transition. They emphasise that it would be illusory to consider the technology a miracle solution to current environmental challenges. AI should instead be considered one tool among others, the use of which must be carefully evaluated in light of its environmental footprint.
Conclusion
The evolution towards a more responsible digital world requires collective awareness and concerted action. The entire digital ecosystem must be transformed beyond technical improvements and innovations in infrastructure management.
This transformation requires close collaboration between the companies developing and deploying the solutions, the public authorities establishing the regulatory framework, and the users whose practices must evolve. Only through this systemic approach can a digital model that is truly compatible with contemporary environmental issues emerge.
