Compas is a popular risk assessment tool used by several states in the United States of America as a piece of information for criminal sentencing. We discuss its legitimacy in the judicial framework. As it relies on machine learning, which use big data to optimize a predictive model, specific questions of fairness arise. It has a cost in terms of transparency and bias might come from the racial bias of the past and current society that generates the training data. We study these questions and show that awareness is welcome to avoid the amplification of norms based on statistics, as it might not be compatible with human notion of justice.

Ask for full text (French).