On some issues related to the fairness of algorithms

International audienceFairness of algorithms is the subject of a large body of literature, of guides , computer codes and tools. Machine Learning and AI algorithms commonly used to accept loan applications, select responses to job offers, etc. are often accused of discriminating against groups.We will begin by examining the relationship between fairness, explainability, and interpretability. One might think that it is better to understand how an algorithm works in order to know whether it is fair, but in fact this is not the case, because transparency or explainability are relative to the algorithm, whereas fairness concerns its differential application to groups of individuals. There is a wide variety of often incompatible measures of fairness. Moreover, questions of robustness and precision are often ignored. The choice of a measure is not only a matter of statistical considerations, but of ethical choices. The "biases" of the algorithms are often only the reproduction of those of previous decisions found in the training data. But they are not the only ones. We will attempt to draw up a typology of the main biases: statistical, societal, cognitive, etc. and to discuss the links with causal models