“Fairness Through Awareness,” by Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel, appeared in 2011 and is one of the foundational formal treatments of fairness in classification. It set out to give mathematical definitions for a problem that had mostly been discussed informally: what does it mean for an automated decision system to be fair?
The paper’s main proposal is individual fairness, captured by the principle that similar individuals should be treated similarly. To make that precise it assumes a task-specific distance metric that measures how alike two people are for the decision at hand, and it requires that people who are close under that metric receive similar outcomes. The hard part, the authors are candid about, is agreeing on the metric itself.
The paper also formalizes group fairness as statistical parity, also called demographic parity, the requirement that the proportion of people receiving a positive decision be the same across protected groups. It analyzes the tension between individual and group notions and connects the whole framework to differential privacy, showing the two share mathematical structure.
Why a business reader should care: this is where the vocabulary that now dominates AI governance, terms like statistical parity and the gap between individual and group fairness, was made rigorous. Anyone setting a fairness requirement for a model is choosing among definitions this paper first laid on the table.