The environmental Cost of Recommender Systems

Abstract

This talk examines the often-overlooked environmental cost of recommender systems, situating modern personalization technologies within the broader context of explosive data growth, information overload, and increasingly energy-intensive machine learning practices. Starting from the societal purpose of recommender systems, it traces how contemporary optimization goals such as engagement, click-through rate, and large-scale deep learning have led to substantial carbon emissions, with evidence showing that a single deep learning recommender systems paper can generate emissions comparable to long-haul intercontinental flights and that, at scale, a year of such research corresponds to tens of thousands of flights and hundreds of thousands of tonnes of CO₂. Beyond research, the talk connects recommender systems to real-world downstream effects, including increased consumption, media overuse, and logistics impacts such as product returns in e-commerce, illustrating how algorithmic choices translate into material environmental costs. The talk concludes by arguing not for abandoning deep learning, but for using it responsibly and purposefully, advocating context-aware recommendation, mitigation strategies, and human-centered design choices that balance user value, business goals, and sustainability consideration