A framework for measuring and reducing the environmental impact of AI

Systems that use artificial intelligence (AI) consume energy and resources to operate. But designing AI that consumes as little as possible runs the risk of a rebound effect: when AI becomes more efficient, less expensive, and easier to implement, its use may increase... and so may its environmental impact! AFNOR has published a document, AFNOR Spec 2314, which reviews the subject and provides keys to avoiding pitfalls and moving towards frugal artificial intelligence.

Frugal AI: 31 best practice sheets

For the working group led by AFNOR, comprising around forty stakeholders led by the General Commission for Sustainable Development (La Poste, Hub France IA, Ademe, EcoInfo, etc.), the concept of frugality involves redefining needs (what is necessary?) and uses (how can AI be used more effectively?). A frugal AI service is therefore a service for which:

• the necessity of using an AI system rather than another less resource-intensive solution to achieve the same objective has been demonstrated;
• best practices are adopted by the producer, supplier, and customer to reduce the environmental impact of the service using an AI algorithm;
• uses and needs aim to remain within planetary boundaries and have been previously questioned.

Frugal AI: communicating without greenwashing

The AFNOR Spec standard sets out a methodology for assessing environmental impacts using a life cycle approach, and provides 31 best practice sheets and recommendations for accurately communicating the frugal nature of an AI service. It is intended for all stakeholders who use or develop AI services and are required to report on their CSR policy, or who must incorporate environmental criteria into their procurement of services that include AI systems, particularly for public procurement.

To stand out in a call for tenders or in the eyes of its customers, an AI supplier or producer may need to publish environmental information about its AI system or service, while clients will need benchmarks to assess the reliability of these statements. The framework therefore recommends that certain information be specified, for example to lend credibility to a quantitative assessment of environmental indicators over the life cycle. The supplier or producer will then specify the scope of the analysis.