Two years after my work on the predictive justice software Compas during my Philosophy Degree, I wanted to summarize my takeaways from this use case, using also my new experience coming from my PhD team that works on fairness as well. Alain Desrosières was a statistician who succeeds in highlighting how the processus of getting numbers used for the statistics and how they are performed embed political decisions. Numbers are built, and thus come from choices which relevance can be discussing by opening the black lox of their creation and analysis.
Even if I believe that the content of his book La politique des grands nombres is integrated to the culture of every good engineer, this book is a seminal work when one wants to deepen the relationship between social impact and data-based learning. This workshop was the opportunity to see how sociologists and some of his former colleague approach his work.
My contribution was on the controversy on racial bias in Compas, and how it has been quantified by the company selling the software, by the newspaper Propublica and by machine learning community on fairness, especially pointing shortcomings on the use of biased dataset for checking bias.