Source-linked AI summary
Fairness in Machine Learning: Lessons from Political Philosophy
Reuben Binns
TL;DR
The paper asks how fairness in machine learning should be operationalised and how discrimination, equality, and justice should be understood. It surveys relevant moral and political philosophy to clarify fair ML debates, concluding that egalitarian considerations may better capture algorithmic wrongness than some traditional discrimination accounts.
Problem
Fair ML lacks a settled operational meaning, while proposed definitions embed competing assumptions about discrimination, fairness, equality, and justice.
Method
The paper reviews moral and political philosophy to clarify conceptual distinctions and situate emerging discrimination-aware and fair ML literature.
Results
The analysis suggests that mental-state accounts of discrimination do not naturally transfer to algorithmic decision-making, while broader egalitarian norms may offer a better foundation.
Takeaways & Limitations
Fairness analysis should examine the underlying justice considerations of an application rather than rely only on narrow, static protected-class categories.
Takeaways & Limitations
Contextually appropriate fairness may depend on factors not typically present in the data available for a machine-learning task.
Abstract
from arXiv · showhide
What does it mean for a machine learning model to be `fair', in terms which can be operationalised? Should fairness consist of ensuring everyone has an equal probability of obtaining some benefit, or should we aim instead to minimise the harms to the least advantaged? Can the relevant ideal be determined by reference to some alternative state of affairs in which a particular social pattern of discrimination does not exist? Various definitions proposed in recent literature make different assumptions about what terms like discrimination and fairness mean and how they can be defined in mathematical terms. Questions of discrimination, egalitarianism and justice are of significant interest to moral and political philosophers, who have expended significant efforts in formalising and defending these central concepts. It is therefore unsurprising that attempts to formalise `fairness' in machine learning contain echoes of these old philosophical debates. This paper draws on existing work in moral and political philosophy in order to elucidate emerging debates about fair machine learning.
1. Introduction
Fair machine learning asks how fairness and non-discrimination can be formalised for operational use, amid competing mathematical measures and longstanding philosophical debates about equality and justice.
- Consequential model-based decisions can reproduce real-world discrimination and unfairly deny protected groups loans, insurance, or employment opportunities.
- Fair ML therefore requires operational definitions of fairness, but proposed measures compare group treatment in materially different ways.
- Statistical or demographic parity compares overall positive or negative classification rates, but can ignore legitimate grounds for differing outcomes.
- Fairness measures can be mathematically incompatible, forcing choices among metrics before technical detection and mitigation proceed.
- The paper uses moral and political philosophy to clarify fair ML’s concepts, map relevant debates, and identify issues for future algorithmic-fairness research.
2. What is discrimination, and what makes it wrong?
The paper examines philosophical accounts of wrongful discrimination and their fit with algorithmic decision-making. It argues that mental-state and individual-treatment accounts face difficulties, motivating broader egalitarian foundations for algorithmic fairness.
- Paradigm discrimination involves differential treatment based on salient social-group membership by decision-makers distributing harms or benefits.
- Mental-state accounts make discriminatory wrongness depend on decision-makers’ animosity, preferences, bad intent, or related attitudes.
- Mental state accounts: Because AI systems do not possess contempt, animosity, or disrespect, mental-state accounts may not naturally classify algorithmic decisions as wrongful discrimination.
- Failing to treat people as individuals: Statistical discrimination uses group-level generalisations to infer individuals’ attributes or future behaviour, potentially reducing firms’ risk when direct evidence is unavailable.
- Failing to treat people as individuals: Treating people differently based on group generalisations is not necessarily wrongful discrimination, because permissibility may depend on errors, affected people, and accuracy-improvement costs.
- These difficulties suggest that broader egalitarian norms may provide a better foundation for theories of algorithmic fairness.
3. Egalitarianism
Egalitarian theories ask what should be equal, when unequal outcomes are acceptable, and how historical and social context should shape fairness judgments in machine learning. These debates imply that fairness metrics must be matched to the relevant currency, social sphere, responsibility structure, and type of harm.
- Egalitarianism treats equal treatment or distribution as central, but philosophers disagree about whether equality directly explains discrimination's wrongness.
- 3.1. The currency of egalitarianism and spheres of justice: Algorithmic outcomes can distribute resources, welfare, capabilities, or political status, so the relevant egalitarian currency may vary by decision context.
- 3.1. The currency of egalitarianism and spheres of justice: Fairness metrics appropriate for economic decisions may differ from those appropriate for civil justice, where parity of outcome can support social solidarity despite unequal base rates.
- 3.2. Luck and desert: Luck egalitarianism permits inequalities arising from informed choices but seeks correction for brute luck, while critics argue that some chosen inequalities still warrant compensation.
- 3.3. Deontic justice: Applying egalitarian principles requires empirical analysis of how inequalities arise, including historical responsibility, social structures, and the causes of unequal base rates.
- 3.4. Distributive versus representative harms: Representational fairness addresses harms to how groups appear in cultural or algorithmic artefacts, where equal representation or ideological weight may matter beyond differential group impacts.
4. Conclusion
Fair machine learning commonly intervenes during data preparation, model learning, or post-processing, but this focus can narrow attention to fixed protected classes and omit context. Philosophical accounts encourage examining why groups are protected and which justice concerns matter, although practical application may be limited by missing information.
- Current fair machine learning approaches typically intervene during data preparation, model learning, or post-processing.
- A narrow focus on prescribed protected classes can overlook why those classes are protected and how they relate to the justice concerns of a specific application.
- Philosophical accounts encourage context-sensitive reflection on which factors matter for fairness and why.
- Fair machine learning may be limited by missing protected characteristics and information about responsibility, culpability, desert, socioeconomic circumstances, and life experience.