Equal Concern and Respect
In the [YOU] on AI Field Guide
The cycle that began with [YOU] on AI holds that the machine amplifies the signal it is fed. Equal concern and respect identifies the signal that must not be fed to a sorting machine in certain domains: the treatment of persons as instances of their demographic categories rather than as the individuals they are. The claim is not that statistical reasoning is always wrong—Dworkin's own account of insurance and risk-spreading distinguishes between tolerable and intolerable uses of group statistics—but that the stakes matter, and that in the domains where liberty, livelihood, housing, and the administration of justice are decided, treating a person as a token of a type is a violation of her standing as an equal, regardless of whether the classification is accurate.
Equal respect also clarifies why the human in the loop is not merely a technical requirement for accountability but a normative necessity. A human reviewer is not there to spot-check the machine's accuracy. She is there to be the locus of the dialogical relationship that equal respect requires: the accountable adversary who can be required to articulate a principle, who can be challenged on the grounds that the principle fails to treat the person as an equal, and who can be held to the consequence of getting it wrong. A system in which the human merely ratifies the model's output has not provided equal respect; it has provided a person-shaped surface over an unaccountable process.
Origin
The formulation was developed most fully in Dworkin's Sovereign Virtue (2000), where he argued that the abstract right to equal concern and respect is the one right so fundamental that it grounds all the others. It does not derive from utility, from contract, or from natural law; it is the foundation on which the legitimacy of liberal political institutions depends, the claim that any legitimate government must satisfy before its other decisions become authoritative. The distinction between concern and respect was present in Taking Rights Seriously (1977) but became the organizing principle only as Dworkin's project matured.
Dworkin distinguished equal concern from equal treatment: equal concern does not require treating people identically, because treating people identically may fail to treat them as equals when their situations are genuinely different. The egalitarian distribution that treats equal concern as requiring equal resources may in fact disadvantage a person with greater needs; the equality that matters is the equal weight given to each person's life, not the equal quantity of resources allocated. This distinction has implications for algorithmic fairness: demographic parity, equalized error rates, and similar metrics may or may not express equal concern, depending on the circumstances. The question cannot be answered by the metric. It requires the normative judgment that Dworkin always insisted belongs outside the optimization loop.
Key Ideas
The Two Dimensions. Equal concern holds that each person's life matters equally and that institutions must not discount the suffering or flourishing of any person on the basis of group membership. Equal respect holds that each person is a responsible agent, the author of her own life, who must be treated as capable of forming and acting on her own conception of the good. An automated system that treats a person purely as a predictable object violates the second dimension without necessarily violating the first: it may assign equal weight to her predicted welfare while denying her the standing as an agent that respect requires.
The Sorting Machine as Denial of Respect. Every sorting system works by classifying persons into types and making decisions based on the type's characteristics. The more accurate the classification, the more firmly the individual is subsumed in the type and the less room there is for the claim “but I am different.” Accuracy is therefore not exculpatory under equal respect; it is sometimes the most precise form of the violation. A perfectly accurate risk score that determines a person's fate based on the statistical behavior of people who resemble her has denied her the one thing equal respect requires: the treatment of her behavior as her own, not as an instance of her category's.
Foothold and Standing. Dworkin argued that equal concern and respect requires that a person always retain a foothold—the door to “but I am different” must remain open. When a decision is made about a person on the basis of group statistics, and she can contest it only by disputing the statistics rather than by showing that her individual case is different, the foothold has been removed. This is the practical disempowerment that equal respect prohibits, and it is also the design logic of the interpretable AI movement: if the model cannot tell you why this individual was denied, it cannot provide the foothold that equal respect requires.
Debates & Critiques
The most pressing debate about equal concern and respect in the AI context is whether the principle requires individualized assessment or merely aggregate fairness—whether it is satisfied by a system that produces equal outcomes across demographic groups, or whether it requires that each individual be addressed in her particularity regardless of the group distribution. Dworkin's framework clearly requires the latter: aggregate fairness statistics measure the system's behavior across populations, not its treatment of persons. But critics argue that individualized assessment is practically impossible in high-volume administrative contexts, and that the realistic choice is between aggregate fairness and no fairness constraint at all. Dworkin's framework does not resolve this practically; it names it as a genuine moral cost of the administrative context rather than a mere technical limitation. A second debate concerns Dworkin's own account of legitimate uses of group statistics in insurance and risk-spreading: he distinguished between uses of statistics that reflect morally arbitrary features of a person's situation (and therefore justify redistribution) and uses that reflect choices for which the person is responsible (and therefore do not). This distinction, applied to algorithmic systems, requires exactly the normative judgment about which features are morally arbitrary and which reflect genuine agency that Dworkin always insisted was outside the range of statistical analysis.
Further Reading
- Ronald Dworkin, Sovereign Virtue: The Theory and Practice of Equality (Harvard University Press, 2000) — the fullest development of equal concern and respect
- Ronald Dworkin, Taking Rights Seriously (Harvard University Press, 1977) — the founding formulation
- Ronald Dworkin, Justice for Hedgehogs (Harvard University Press, 2011) — the final synthesis grounding all values in a unified account of dignity
- Sonja Starr, “Evidence-Based Sentencing and the Scientific Rationalization of Discrimination,” Stanford Law Review 66 (2014)