The Hidden Cost of AI for Science

Abstract

Artificial intelligence is transforming scientific research, enabling new discoveries and increasingly sophisticated methods of analysis. However, these advances come with computational and environmental costs that are often overlooked. In this talk, I explore the energy consumption and carbon footprint of AI research, drawing on empirical studies of machine learning in recommender systems. I discuss how experimental practices, model complexity, and hardware choices influence these costs, and whether more computationally expensive approaches necessarily lead to better results. Finally, I highlight practical ways to reduce the environmental impact of AI research and argue for making resource consumption a more explicit part of how we evaluate scientific progress.