Multi-dimensional discrimination in AI systems

What is it about?

AI-driven decision-making can lead to discrimination based on protected characteristics such as race, gender, or age. Fairness-aware machine learning methods aim to address bias in AI models, but most approaches focus on a single attribute. However, discrimination is often multi-dimensional, with individuals facing bias based on multiple characteristics. This issue, known as "multi-dimensional discrimination", has been less explored in machine learning literature. In this work, we provide an overview of various forms, including cumulative, intersectional, and sequential discrimination. Central to this paper is a comparative evaluation of legal and machine learning approaches to addressing multi-dimensional discrimination, where we draw connections, identify limitations, and highlight open research directions.

Featured Image

Photo by Tingey Injury Law Firm on Unsplash

Photo by Tingey Injury Law Firm on Unsplash

Why is it important?

Human identities are complex and encompass multiple dimensions. Algorithmic discrimination can occur when individuals face bias based on more than one protected characteristic, a challenge that is exacerbated by data scarcity as the number of protected dimensions increases. This can also lead to fairness gerrymandering, where addressing discrimination in each dimension in isolation might inadvertently increase bias against intersectional groups.

Perspectives