<p>At the turn of the millennium, the increased availability of large data sets and computing power led to the rise of machine learning in Artificial Intelligence (AI). While there has been much excitement about AI’s achievements since then, concerns about its drawbacks, including its environmental impact, are growing. This paper explores the possibility of a more sustainable form of AI. It argues that, although such a transition is feasible, it would require AI to abandon its dominant ‘Big Data’ approach and adopt a ‘Small Data’ paradigm. The paper intends to contribute to the interdisciplinary field of AI ethics in three ways. Firstly, it proposes framing the current dominant model of AI as ‘Big Data AI’, understood as a system that is intrinsically voracious in terms of its insatiable demand for (behavioural) data, energy, and natural resources. Secondly, it suggests a comprehensive, multi-factor taxonomy of the environmental impact of AI throughout its entire lifecycle. This taxonomy considers the environmental implications of AI in terms of rare earth mining, water consumption, energy usage, carbon footprint and pollution during the manufacturing, development, deployment and disposal phases of the AI lifecycle. Finally, it explores the notion of Small Data as a potentially more environmentally sustainable alternative to Big Data AI and provides an overview of Machine Learning (ML) techniques able to deal with data or computing power scarcity that have been suggested in AI literature. It concludes that Small Data techniques have the potential to reduce the environmental impact of AI, particularly its CO₂ footprint, water consumption and energy usage, but only if embedded within a broader sociotechnical process involving political action and economic transformation. </p>