Algorithmic fairness in the financial sector

To address this challenge, the long-dominant approach known as “fairness-through-unawareness” has consisted in excluding sensitive variables from statistical processing. While this method has always been subject to debate, it has now been rendered largely obsolete by the development of artificial intelligence (AI) models, which are capable of reconstructing the information contained in these variables. More broadly, the rapid growth of AI is reshaping how issues of algorithmic fairness are addressed, making it necessary to develop approaches that are more explicit, more robust and better suited to the complexity of current systems.

Against this backdrop, this discussion paper first sets out the legal framework for algorithmic fairness in the financial sector. It shows that although non-discrimination is a firmly established principle, its implementation is more complex when decisions are based on statistical models, and even more so on AI systems. The European Regulation on Artificial Intelligence (the “AI Act”) adds to this framework by setting out fairness requirements for so-called “high-risk” AI systems and by reaffirming a principle of non-discrimination for all AI systems. In the financial sector, this regulation dovetails with the rules on customer protection in the banking and insurance industries, which also set out requirements for fairness, often based on a principle of “protection through abstention” (from granting or selling), whereas the AI Act places greater emphasis on the risks of exclusion. A comparison between the legal frameworks in other countries then highlights the diversity of approaches.