The representative individuals approach to fair machine learning - Original research

Aug 3, 2026

Upload by Levi

Fair machine learning is commonly framed probabilistically, even though many real decision systems behave deterministically for individuals, yielding effectively zero-or-one outcomes. The paper explains the rationale behind this probabilistic language by interpreting fairness measures’ statistical reference classes as specifying the probability that hypothetical people, representative of social roles, receive certain goods. These “representative individuals” ground what is owed to actual individuals whose individual chances may be extreme. The work argues this framework offers significant advantages over competing interpretations in fair ML.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Introduction

Hiring algorithm and uneven errors

The intuitive defense

Problems with the intuitive defense

Alternative approaches and the representative individuals approach

AI and Ethics [https://doi.org/10.1007/s43681-025-00675-y](https://doi.org/10.1007/s43681-025-00675-y) ORIGINAL RESEARCH The representative individuals approach to fair machine learning Clinton Castro1 · Michele Loi2 Received: 27 August 2024 / Accepted: 4 January 2025 © The Author(s) 2025 Abstract The demands of fair machine learning are often expressed in probabilistic terms. Yet, most of the systems of concern are deterministic in the sense that whether a given subject will receive a given score on the basis of their traits is, for all intents and purposes, either zero or one. What, then, can justify this probabilistic talk? We argue that the statistical reference classes used in fairness measures can be understood as defining the probability that hypothetical persons, who are representative of social roles, will receive certain goods. We call these hypothetical persons “representative individuals.” We claim that what we owe to actual, concrete individuals—whose individual chances of receiving the good in the system might be extreme (i.e., either zero or one)—is that their representative individual has an appropriate probability of receiving the good in question. While less immediately intuitive than other approaches, we argue that the representative individual approach has important advantages over other ways of making sense of this probabilistic talk in the context offair machine learning. Keywords Fair machine learning · Algorithmic bias · Fairness · Technology ethics · Philosophy of technology 1 Introduction Consider Hiring Algorithm.1 A machine learning system is used to determine which applicants should get a firstround interview. Unfortunately, the machine distributes its errors unevenly: it identifies men who are qualified for the job as worthy of interviews at much higher rates than it does women who are qualified. It would be very natural to say in this case that the machine is biased in virtue of its differential error rate2 across groups and that further, in virtue of this, it is unfair. 1 This is a fictionalization based on a real case; see Dastin (2018) for the real-world case. 2 It is inessential to this example that it focuses on a case where error rates (as opposed to predictive power) is different across groups. This 􀀍 Clinton Castro [clinton.g.m.castro@gmail.com](clinton.g.m.castro@gmail.com) 1 University of Wisconsin-Madison, Madison, USA 2 Politecnico di Milano, Milan, Italy This thought aligns with how many—including ourselves—are inclined to think about fairness in the context of machine learning. But how, exactly, are certain grouplevel asymmetries (such as unequal error rates3) and unfairness connected to one another when it comes to machine learning? Here is one intuitive defense recently made explicit by Sune Holm [1]: Intuitive defense.4 (1) When distributing goods (e.g., callbacks), it is important from the perspective of fairness that paper does not take a stand as to which of those different—and often incompatible Chouldechova [8]—group-level ratios are preferable. 3 We are not engaging in questions of which fairness measures apply in which contexts in this paper. For discussion of these issues, see, Hellman [9], Hedden [10], Long [11], Holm [1], Grant [12], and Loi et al. [13] . 4 This articulation of Holm [1] bears some semblance to the articulation given by Castro and Loi [14] . While “the intuitive defense”mirrors Holm—who, we should, mention is channeling Broome [3] in (1)—we take it that the intuitive defense at least roughly represents a fairly widespread and natural thought: Fairness is about giving individuals fair chances, and statistics (such as error rates) give us a glimpse into individual chances. 1 3 individuals with similar claims to the good have similar chances of receiving the good. (2) We can assess whether (1) is met by considering whether the appropriate group-level ratio (e.g., equal error rates) is in proper proportio

FAQs

Why does fairness discussion use probabilistic talk when ML systems often yield deterministic outcomes per individual?

The paper argues that fairness measures use statistical reference classes that can be understood as defining probabilities for hypothetical people representative of social roles, which makes the probabilistic framing conceptually justified.

What are the two key problems the paper identifies with the intuitive defense of fair ML?

It identifies the equal probabilities talk problem (how group ratios relate to individual-level probabilities) and the narrow reference class problem (how individuals are grouped into reference classes).

What is the representative individuals approach, and how does it relate to fairness obligations?

The approach interprets fairness reference classes as probabilities for “representative individuals,” and claims what we owe to actual individuals is that their representative individual has an appropriate probability of receiving the relevant good.

Aug 3, 2026

Upload by Levi