What is the conflict between different definitions of fairness?

The conflict between different definitions of fairness in machine learning and AI arises because various metrics and approaches to fairness often contradict each other and cannot be simultaneously satisfied.

For example, the main types of fairness include:

• Demographic parity: equal distribution of positive outcomes across groups.
• Equalized odds: equal error rates (false positives and false negatives) between groups.
• Predictive parity: equal accuracy of predictions for different groups.

The problem is that with different data distributions and baseline characteristics of groups, it is impossible to satisfy all these criteria at once. This leads to the need to choose which definition of fairness is more important in a specific context, raising ethical and technical dilemmas.

Thus, the conflict lies in the fact that different definitions of fairness reflect different values and goals, and achieving them simultaneously is often impossible, requiring compromises and transparency in decision-making.