Used to show no proposed fairness criteria besides CWG are genuine requirements of fairness by showing a perfectly fair algorithm violates all fairness criteria.

Thought Experiment

  • 40 individuals each given a coin with some bias probability of landing heads between 0 and 1. Each coin has a bias written on it
  • Individuals randomly assigned two rooms, Room A or B
  • Aim is to predict for each individual, if their coin will land heads or tails and assign them a score representing this likelihood
  • Assume we use the algorithm that we only predict heads if the number of the coin is greater than 0.5
  • Suppose random assignment produces following result:
Heads ProbaabilityRoom ARoom B
75%12
12.5%8
60%10
40%10
  • This violates every fairness criteria except CWG because A’s have more extreme probabilities than Bs