Today’s technologies are becoming more and more pervasive and advanced software systems can replace human beings in many different tasks. This is especially true in the case of automated decision-making systems based on machine learning (ML). Important ethical implications arise when such decision systems are used in sensitive contexts (e.g., justice or loans). The elicitation of these implications, that is, of the ethical requirements behind ML-based systems is a new challenge we must address to avoid societal risks. This is particularly urgent for fairness since this notion lacks a precise and commonly accepted definition, thus hampering its assessment. This paper aims to give a comprehensive definition of fairness, present a unified taxonomy of alternative interpretations, define a new decision tree that can guide the choice of the correct interpretation, and carry out a preliminary assessment with experiments in a real-world context.