"Fairness" in AI sounds simple but is mathematically complex — different definitions of fairness often conflict with each other.

Competing Definitions of Fairness

• Demographic parity: Equal positive outcome rates across groups
• Equal opportunity: Equal true positive rates across groups
• Calibration: Predicted probabilities match actual rates for all groups
• Individual fairness: Similar individuals treated similarly

The Impossibility Result

Mathematically, it's impossible to satisfy all fairness definitions simultaneously. Every AI system embeds a value judgment about which definition of fairness matters most — and that choice should be made explicitly, not by default.

Reference:

Fairness and Machine Learning textbook

https://fairmlbook.org/