Afnor Spec IA frugale

Summary

This Afnor Spec aims to provide an initial operational framework for assessing the environmental impact of an artificial intelligence system and service, as well as ways to promote frugal AI services for all organizations, both public and private.

The objective of this document is to share a guide outlining definitions, methods, and best practices useful for evaluating and reducing the environmental impact of an AI system and the digital services that use it.

The framework successively covers:

1. contextualization issues (definitions, life cycle, usage framework)

2. evaluation methodology

3. the 31 best practices to implement

4. how to communicate about the project

Within the framework, a key constraint for initiating a frugal AI approach is: “Frugality must be respected in all aspects of a service, such as its objectives (are the service’s objectives compatible with planetary boundaries?)” A key constraint for communicating about a project claiming to adhere to frugal AI principles comes from various recommendations on the subject, summarized on page 50 of the Afnor Spec.

The editorial perspective

The Afnor Spec Frugal AI is a highly relevant project that sets a strong direction for what frugal AI should be. The document establishes an initial foundation for standardizing discussions and projects that claim to follow the principles of frugal AI. It is well-structured, readable, and accessible to anyone looking to develop an artificial intelligence-based project.

From the beginning, the document does not shy away from addressing the purpose of a digital service. In this regard, the Afnor Spec Frugal AI is cautious about the framework within which this reference document can be used and the legitimacy of claiming adherence to it. Various safeguards embedded throughout the Afnor Spec naturally limit how the framework can be applied. However, the guidelines could be clearer and more directly formulated to better constrain the purpose of services that claim to follow frugal AI principles.

The "Communication" section lays the groundwork for how to promote a frugal AI approach while imposing strict guidelines on stakeholders to minimize the risk of greenwashing. The summary table on page 50 serves as a must-have checklist for many projects—not just for frugal AI initiatives but also for digital eco-design projects. For example, one of the key guidelines for communication requires assessing the "potential impact shifts of a service".

One downside is that nearly a quarter of the best practices mention the term 'frugal' or 'frugality'—defining frugal AI solely based on a lexicon centered around the word 'frugal' does not always provide the desired level of precision.

Another limitation is the formal structure of the Spec, which distances readers from best practices starting from page 62 (out of 102 pages). The evaluation section is given significant emphasis before that, reinforcing the importance of measuring impact and redirecting users to a standardized methodology—specifically, the one established by Ademe through the PCR Digital Services. Some redundancies between the Spec and PCR could have been avoided to streamline the document and reduce duplication, which increases the risk of obsolescence when the PCR is updated.

That said, one of the most relevant aspects of this section is the work on evaluating indirect impacts related to a service’s usage. This approach helps service designers adopt a more systemic view of environmental impacts.

To fully understand the 31 best practices and enhance their usability, it is recommended to download the .csv version available on data.gouv.fr.