The environmental impact of AI: a framework for energy consumption in machine learning services

Purpose This study aims to develop a framework to analyze the energy consumption of machine learning (ML) services. It enables a process-oriented understanding of the environmental impacts and identifies actionable levers for their mitigation. Indeed, ML services are increasingly integrated into organizational processes generating rising energy demand and environmental impact, highlighting the need for solutions to assess these issues. Design/methodology/approach Using a design sience research approach, first an initial framework drawing on Cross Industry Standard Process for Data Mining (CRISP-DM) was developed through a targeted literature review, and observations of different recommender systems. Then, a two-step assessment (pre-assessment and expert evaluation involving academic and practitioner stakeholders) was conducted. Insights from these evaluations enriched the final framework proposal. Findings This research reveals that along the CRISP-DM process different factors impact energy consumption and must be considered. Three different phases and related energy relevant aspects were named. The evaluation of the framework as an artifact underlines its statement and value for research and practice. Research limitations/implications The framework provides a holistic overview of relevant factors that must be considered in the context of energy consumption of ML-based artificial intelligence (AI) services. Limitations can be found in the neglected model accuracy, the selected sources and the applied research approach itself. Practical implications The developed framework can be used by practitioners to understand and evaluate important factors of the energy demand that should be considered before and when building a ML-based AI service in practice. Originality/value This paper develops and presents a framework for the environmental impact of AI, focusing on ML approaches. The framework provides a holistic view of energy consumption in ML services based on the previous literature and recommender systems. In addition, it is process-based and enables its usage at different milestones within a project.