Abstract

Detecting bias in AI algorithms has become a necessity since the adoption of the European AI law. In particular, online sales platforms enable consumers to find different product offers from various sellers. The operators of these platforms use complex algorithms to determine which sellers and products consumers see first on their sites. These “recommendations” can benefit consumers by facilitating their search, but they also have a huge impact on sales. Through an audit of one such algorithm, we show that these algorithms can be biased. We will explain the link between global and local measurements, and study the impact of these deviations. We will attempt to quantify the impact of these biases using counterfactual generation methods based on optimal transport.

Biography

Jean-Michel Loubes is currently Director of Research at Inria. He also holds a research chair at the ANITI AI Cluster in Toulouse. His research focuses on the robustness and biases of machine learning algorithms. As such, he is a member of the AFNOR commission responsible for drafting European AI Act standards.