Legal Taxonomies of Machine Bias: Revisiting Direct Discrimination

Previous literature on ‘fair’ machine learning has appealed to legal frameworks of discrimination law to motivate a variety of discrimination and fairness metrics and de-biasing measures. Such work typically applies the US doctrine of disparate impact rather than the alternative of disparate treatment, and scholars of EU law have largely followed along similar lines, addressing algorithmic bias as a form of indirect rather than direct discrimination. In recent work, we have argued that such focus is unduly narrow in the context of European law: certain forms of algorithmic bias will constitute direct discrimination [1]. In this paper, we explore the ramifications of this argument for existing taxonomies of machine bias and algorithmic fairness, how existing fairness metrics might need to be adapted, and potentially new measures may need to be introduced. We outline how the mappings between fairness measures and discrimination definitions implied hitherto may need to be revised and revisited.