Attesting digital discrimination using norms

Abstract

More and more decisions are delegated to Machine Learning (ML) and automatic decision systems recently. Despite initial misconceptions considering these systems unbiased and fair, recent cases such as racist algorithms being used to inform parole decisions in the US, low-income neighborhood's targeted with high-interest loans and low credit scores, and women being undervalued by online marketing, fueled public distrust in machine learning. This poses a significant challenge to the adoption of ML by companies or public sector organisations, despite ML having the potential to lead to significant reductions in cost and more efficient decisions, and is motivating research in the area of algorithmic fairness and fair ML. Much of that research is aimed at providing detailed statistics, metrics and algorithms which are difficult to interpret and use by someone without technical skills. This paper tries to bridge the gap between lay users and fairness metrics by using simpler notions and concepts to represent and reason about digital discrimination. In particular, we use norms as an abstraction to communicate situations that may lead to algorithms committing discrimination. In particular, we formalise non-discrimination norms in the context of ML systems and propose an algorithm to attest whether ML systems violate these norms.