Where does fairness collide?

In the study of behaviour at work, fairness is a perception an employee holds: a judgment about whether outcomes, procedures, and treatment on the job were just. Colquitt's 2001 On the dimensionality of organizational justice validated a measure that splits this judgment into separate strands, showing that people weigh the fairness of what they get, of how decisions are made, and of how they are treated as distinct things. Here fairness is a measured attitude that predicts trust, commitment, and behaviour, so the fuller anatomy of these workplace strands lives at organizational justice. This referent is a rated feeling inside an organization, not a choice in a bargaining experiment, a formal property of a classifier, or a principle for the basic structure of society.[1]

Unit

A worker's rated judgment that a workplace outcome, procedure, or treatment was fair

Recovery means

Not applicable

Measured by

Captured by validated survey scales that ask employees to rate the fairness of outcomes, procedures, and interpersonal treatment

In behavioral economics, fairness names a preference that shows up in choices, not a survey rating: people will pay a real cost to punish an allocation they find unfair. Güth, Schmittberger, and Schwarze's 1982 An experimental analysis of ultimatum bargaining had one player propose how to split a sum and a second player accept or reject, where rejection left both with nothing. Responders often turned down low offers, giving up money to refuse a split they judged unfair, which a purely self-interested model does not predict. This referent is a behavioural taste inferred from what people actually reject, distinct from an attitude scale at work, a classifier criterion, or a philosophical principle.[2]

Unit

A revealed willingness to forgo payoff in order to reject an allocation judged unfair

Recovery means

Not applicable

Measured by

Inferred from choices in bargaining and allocation experiments, such as the rate at which responders reject low offers

In machine learning, fairness is not a feeling or a preference but a precise statistical property required of a model's predictions across groups, such as equal error rates or equally calibrated scores. Kleinberg, Mullainathan, and Raghavan's 2017 Inherent trade-offs in the fair determination of risk scores proved that several natural criteria of this kind cannot all hold at once except in degenerate cases, so a designer must choose which definition of fair to satisfy. Here fairness is a formal constraint an algorithm either meets or violates, defined by equations on predicted and true labels. This referent is a mathematical condition on a classifier, unlike a rated workplace attitude, an experimental preference, or a principle of justice.[3]

Unit

A formal condition (for example equal false-positive rates or equal calibration) that a model's predictions satisfy across groups

Recovery means

Not applicable

Measured by

Checked by computing group-wise statistics of a classifier, such as calibration, false-positive rates, or false-negative rates

In moral and political philosophy, fairness names a normative ideal for how a society should be arranged, argued for rather than measured. Rawls's 1971 A Theory of Justice set out justice as fairness, deriving principles that free and equal people would choose for the basic structure if they reasoned behind a veil of ignorance that hid their own position. The result is a claim about which arrangements are just, defended by argument about what such people could accept. This referent is a principle for institutions, worked out by reasoning about what people would agree to, not a workplace perception, an experimental choice, or a statistical property of a model.[4]

Unit

A defended principle stating which social arrangements count as just

Recovery means

Not applicable

Measured by

Not measured empirically; argued for by reasoning, for example from a hypothetical original position