Contribution
Burrell argues that ML-ethics research leans heavily toward distributive, computationally tractable definitions of fairness, and she explains the skew sociologically, through how computer science reproduces itself as a field. She holds that allocative fairness is an inherently limited account of justice. She proposes instead Iris Marion Young’s account of injustice as domination and oppression, which she argues fits the algorithmic harms the community already acknowledges better. On that account, legislation, participatory assessment, user repurposing and resistance, and activism for outright bans all count as remedies alongside technical fixes.
Skim
Abstract, figures and selected sections, read when the work was seeded. Not enough to state its assumptions or results exactly; a Read note replaces this one.
- §1–§2: the author reads the AAAI 2020 “godfathers” panel (“the field can fix its flaws”) as an instance of disciplinary conservatism. Citing bibliometric studies of FAccT and AIES output, she says distributive-fairness work dominates.
- §2.1: fairness succeeded because it is “efficient” framing that translates harm into CS’s problem-solving idiom. Buolamwini & Gebru (2018) is her example of framing that works. The Google “gorillas” image-labelling case, she argues, is a poor fit: its harm is representational subordination (Collins’s “controlling images”, Noble), not misallocation.
- §2.2: political economy. She describes corporate funding of US CS departments and public-private procurement as advantages for technical-fix work at interdisciplinary venues.
- §3: she summarises the technical and critical objections to formal fairness (Kleinberg et al. 2017 impossibility; Green & Viljoen; Powles’s “seductive diversion”). She contrasts Anderson’s “democratic equality” with Young’s rejection of the distributive paradigm, where justice is “a matter of doing”, with participation and voice at its centre. Her examples are the COMPAS parole case (Glenn Rodriguez) and the Arkansas Medicaid allocation algorithm.
- §3.1–§3.2: the widened solution space is sociotechnical design and empowered intermediaries, strategic-classification work reframed as legitimate recourse, “learning to defer”, facial-recognition bans, protest, and artists withholding work from generative-AI training data.
- §4: defining justice by what can be formalised “gets the proper formulation backwards”. She argues for augmenting human agency rather than replacing it.
Open questions
- It is a counterweight to fairness-metric practice. If the anthology holds practices about fairness evaluation, this is the source that says what those metrics cannot measure.
- It argues from secondary literature and theory and has no new empirical data. The claim that fairness work is “overproduced” rests on the cited bibliometrics (Birhane et al. 2022a and others), which a deeper reading should check.
- It is a position paper, so it probably belongs in the anthology as a note (LIT) with a critical stance rather than as the source of a practice.