Impossibility of fairness

The impossibility of fairness refers to a family of mathematical results demonstrating that no algorithmic decision system can simultaneously satisfy multiple intuitively plausible fairness criteria when base rates differ across protected groups. The landmark results were proved independently by Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan in 2016, and by Alexandra Chouldechova in 2017.

The core theorem shows that demographic parity, equalized odds, and calibration are mutually exclusive in any non-trivial setting where the prevalence of the predicted outcome differs across groups. A system designer must choose which criterion to violate. The choice is not a technical optimization problem. It is a normative decision about which kind of unfairness is more acceptable.

The impossibility results are sometimes misread as showing that fairness is impossible. This is not what the theorems prove. They prove that fairness cannot be encoded as a single mathematical constraint. Fairness requires choosing among incompatible desiderata, and the choice must be justified by ethical reasoning, not by algorithmic optimization.

The systems-theoretic significance of the impossibility results is that they formalize a boundary condition. In any system where classification is applied to groups with different historical experiences, the classification itself must encode a value judgment about which groups should bear which kinds of error. The mathematics does not resolve the judgment. It makes the judgment visible.