This article applies the three parts of Aristotle’s theory of justice (Book V) to algorithmic decision systems, turning “fairness” from a technical parameter into a set of governance requirements.
Distributive justice: the fairness metric is a political choice #
Machine learning research has shown that common fairness definitions (demographic parity, equalized error rates, calibration) are mutually incompatible in realistic conditions — a system cannot satisfy all of them at once. Aristotle diagnosed this 2,300 years early: everyone accepts that justice is proportional to merit, and the real dispute is over the criterion of merit (1131a).
Choosing a fairness metric is therefore choosing a theory of desert. That choice:
- must not be made silently by a vendor’s default setting;
- belongs to accountable public deliberation — it is exactly the kind of question legislatures, rulemaking, and public comment exist for;
- must be documented: which definition was chosen, which were rejected, and why, in language citizens can read.
Corrective justice: contestability and remedy #
Distribution rules, however well chosen, will wrong individuals. Corrective justice (1131b–1132b) demands rectification, which translates into four requirements for any consequential system:
- Notice — the person knows an automated system was involved and what it concluded.
- Reasons — an explanation specific enough to contest.
- Appeal — review by a human with genuine authority to decide otherwise, on deadlines proportionate to the stakes (a benefits cutoff cannot wait ninety days).
- Restoration — when the system erred: correction, back payment where relevant, and repair of downstream records that propagated the error.
Equity: the override power #
Aristotle’s epieikeia (Book V, chapter 10) corrects the law where its universal language fails the particular case — like the builders’ flexible rule on Lesbos that bends to fit the stone. Every algorithm will eventually meet a case its designers never imagined. Equity requires:
- a designated human role with authority to depart from the system’s output in atypical cases;
- logged, reviewable overrides — so equity remains principled correction, not caprice;
- feedback from override patterns into system revision (frequent overrides in one case type mean the universal needs rewriting).
Quick test for any deployed system #
Ask three questions. Who chose the fairness criterion, and where is that recorded? What happens, step by step, when the system wrongs someone? Who can bend the rule when the rule does not fit — and how would we know they did? A system without good answers to all three fails Aristotelian justice regardless of its accuracy statistics.
