Environmental sustainability of Artificial Intelligence: A GUM-based, user-centered measurement framework

Given the increasing adoption of artificial intelligence (AI) across a broad spectrum of applications, along with the urgent need for sustainable development, understanding the environmental sustainability of AI pipelines has become increasingly relevant. In this regard, however, the current state of the art lacks a reliable methodology for measuring environmental sustainability from a user-centered perspective (i.e., by considering all the operations typically performed by end users), which is essential for achieving awareness of the actual sustainability in the development and adoption of AI models. Starting from these considerations, this paper employs a rigorous methodology based 1) on the ISO standards for sustainability assessment and 2) on the Guide to the Expression of Uncertainty in Measurement (GUM) to measure and aggregate the Carbon Footprint required by each stage of an AI pipeline. To the best of the authors’ knowledge, this work constitutes the first to integrate GUM-based approaches into a user-driven AI pipeline. To illustrate the methodology, a case study on the use of AI models in biosignal processing is presented. Without losing generality, the results provide useful insights for implementing more sustainable AI practices, enabling a reliable, environment-oriented assessment of AI pipelines and guiding decisions toward reduced environmental impact.