REVIEW 3 major objections 4 minor 3 references
What is the Point of Fairness? Disability, AI and The Complexity of Justice
T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read The paper argues that a fairness framing of AI ethics for disabled people can reinforce medical gatekeeping and surveillance, and that the field should instead center justice.
desk verdict A useful position paper that concretely applies the fairness-to-justice critique to disability and computer vision, with a few soft spots but a sound core. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The mechanism carrying the argument is the paired case-study comparison of two ethical lenses. Each technology is first seen through a fairness lens—what biases in data or outcomes could be corrected—and then through a justice lens—who gains diagnostic or interpretive authority, what social structures are legitimized, and which disabled people are served or harmed. The fairness-justice distinction, drawn from prior critiques that fairness is modeled on anti-discrimination law and leaves structural oppression untouched, supplies the category that lets the paper show how a technology can be fair yet unjust.
What would settle it
If a longitudinal deployment study of a fairness-adjusted autism-diagnosis aid showed that multiply marginalized children gained access to self-directed support without diagnostic over-reliance, and that sight-assist users overrode automated interpretations in practice with no surveillance spillover, the paper's central claim would be weakened. The study would need to observe actual use, not lab accuracy.
Extended reading notes
Core claim
The central claim is that 'fairness' as usually operationalized in AI ethics—equal treatment for similar cases separated only by disability status—is not merely incomplete but can be harmful when applied to disabled people. In the diagnosis case, a fairness fix such as diversifying training data would leave untouched the question of why diagnosis is the only legitimate route to autistic existence, strengthening the gatekeeping authority of psychiatrists and now algorithms. In the sight-assistance case, even a vision system that treats all users equally can transfer judgment from the user to black-boxed technology and normalize surveillance. The paper concludes that researchers should move beyond fairness and center justice, understood as attention to structural oppression, power, and multiply marginalized disabled people.
Load-bearing premise
The argument depends on the premise that the harms it describes—medical gatekeeping from automated autism diagnosis and epistemic transfer plus surveillance normalization in sight-assistance tools—would actually arise and would not be prevented by any fairness-based remedy.
Editorial extensions
If this is right
- Fairness metrics such as equalized odds or dataset diversification will not by themselves prevent the harms described; a fair autism-diagnosis system can still reinforce medical gatekeeping, and a fair sight-assist system can still defer judgment to black-boxed technology.
- Researchers and designers evaluating assistive AI must ask who holds power over diagnosis and interpretation, not only whether outcomes are equal.
- Design and evaluation should center multiply marginalized disabled people, since benefits to otherwise-privileged disabled people can coincide with harm to those marginalized in multiple ways.
- AI ethics conversations around disability should draw on disability studies and treat technology as part of the construction of disability rather than a neutral add-on.
- Justice-oriented design would question the premises of the technology itself, such as whether early diagnosis is a good outcome or whether vision must be the privileged sense.
Reading between the lines
- A testable extension: if the argument is right, deployed assistive AI optimized only for fairness will show harms that fairness audits miss—such as diagnostic over-reliance or user deference—in longitudinal user studies, while participatory designs that give users veto power over automated interpretations should not.
- The fairness-to-justice shift implies that ethical review of assistive AI should include governance models, not just performance metrics; this could be operationalized as requiring documented consent and user veto over automated outputs.
- Neighbouring domains—employment screening, predictive policing, and health-care allocation—face the same limitation the paper names for disability, so the argument generalizes to any group whose marginalization is structural rather than merely statistical.
- An implicit consequence is that 'fairness' in AI may keep its use for narrow auditing, while 'justice' becomes the standard for whether a technology should exist at all.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper argues that the dominant 'fairness' framing in AI ethics—understood as statistical outcome parity and dataset representativeness—is insufficient for addressing the harms of AI for disabled people, because it ignores structural oppression, epistemic hierarchies, and the social meaning of technologies. Through two computer-vision case studies (automated autism diagnosis and sight-assistance tools), the authors claim that even fairness-improved systems would still reinforce medical gatekeeping, shift judgment away from disabled users, and legitimize surveillance. They conclude by urging a move 'towards notions of justice,' drawing on disability studies and critical data science.
Significance. If the argument holds, the paper makes a timely and important contribution by bringing disability studies to bear on mainstream AI ethics and by challenging the adequacy of fairness metrics as the primary ethical lens. Its strengths include a clear presentation of two concrete case studies, explicit engagement with foundational critical literature (Hoffmann, Conrad, Waltz, Mankoff et al.), and a pointed call to center multiply marginalized disabled people in AI design. The paper is coherent and argued in good faith; the case studies effectively illustrate harms (medical gatekeeping, epistemic transfer, surveillance normalization) that are often invisible in metric-driven fairness discussions. For a position paper in this area, the conceptual contribution is valuable and would help reorient assistive-technology research and AI ethics conversations.
major comments (3)
- [Case Studies, 'AI For Diagnosis' and 'AI For Sight'] The central shift from fairness to justice rests on the claim that the identified harms would persist even under fair deployment—e.g., 'computer vision, even deployed fairly, cements vision as a superior sense and legitimizes surveillance'—and that the 'issue is not one of discrimination against the patient for being autistic but for being a patient.' This persistence claim is load-bearing: if a robust fairness process (e.g., equalized odds across intersectional groups, participatory dataset audits, or co-design with disabled stakeholders) could surface and mitigate the same concerns, the fairness-to-justice dichotomy would collapse into a call for better fairness. The paper asserts this persistence through cited critical theory rather than analyzing why specific fairness remedies would fail. I recommend that the authors either engage with the broader fairness literature—including processual and structural approaches—or qualify the claim to 'metric-based fairness as typically practiced in ML,' which would make the argument more defensible.
- [Conclusion] The paper calls for a reframing 'towards notions of justice,' but the concept of justice is never defined or operationalized. The reader is not told what a justice-oriented design or evaluation would look like in either case study, beyond vague references to 'power' and 'wider social context.' Since the paper's title and central recommendation hinge on this dichotomy, the positive framework is underspecified. A concrete sketch—e.g., drawing on disability-justice principles, or showing how the two case studies would be redesigned under a justice lens—would make the contribution far more actionable and would allow readers to assess whether the proposed shift is genuinely distinct from a more robust fairness framework.
- [Fairness: Defined and Contested] The paper scopes fairness narrowly to statistical outcome parity and data representativeness, following Trewin and Hoffmann. This is a defensible reading of the ML fairness literature, but it risks setting up a strawman: several fairness frameworks in the broader literature address structural concerns, such as fairness as non-domination or fairness as accountability, and are not reducible to parity metrics. The paper would be strengthened by acknowledging these positions and explaining why they too would fail to address the case-study harms, or by explicitly confining the critique to a labeled subset (e.g., 'metric-based individual fairness'). Without this, the dichotomy between fairness and justice appears sharper than it actually is.
minor comments (4)
- [Throughout] The citation numbering is seriously inconsistent: for example, Trewin's 'AI Fairness for People with Disabilities' is cited as [27] in the introduction but appears as [29] in the Fairness section; Mitzi Waltz's 'Autism=death' is cited as [35] in the text, but the reference list has it as [32]; and the World Institute on Disability is cited as [32] in the introduction but [34] later in the text. These mismatches make the bibliography difficult to verify and must be corrected.
- [Introduction] 'we would be remis to deny' should be 'remiss'.
- [AI For 'Sight', Concerns Raised Through a Justice Lens] The claim that assistive seeing tools 'legitimize surveillance' is supported by a plausible mechanism (the AT Effect, the savior narrative), but the sentence 'who's to say blind people aren't among the users of policing technologies' is speculative and could be strengthened with an example or a reference showing actual integration or misuse of similar technologies.
- [General formatting] A large copyright/license boilerplate appears in the body of the manuscript; this should be removed from the submission version. Heading styles are also inconsistent (e.g., 'Case Studies of AI in Assistive Technology' appears in a different font than other headings).
Circularity Check
No circular derivation found: the paper's fairness-to-justice argument is grounded in external critical and empirical literature, and its self-citations function as supporting examples rather than load-bearing premises.
full rationale
This is a conceptual and critical paper, not a derivation with fitted parameters or mathematical predictions. The paper defines fairness as the statistical 'similar cases, same outcome' framing, but the argument that this framing is inadequate is sourced to Hoffmann's external critique and to disability-studies scholarship rather than being entailed by the paper's own definition. The two case studies are illustrative: the claim that harms persist even under fair deployment (e.g., 'computer vision, even deployed fairly, cements vision as a superior sense and legitimizes surveillance') is an interpretive argument supported by cited prior work in medicalization, disability studies, and empirical studies of blind users' experiences, not a quantity fitted from one subset of data and then 'predicted' as a result. The self-citations (Keyes 2018; MacLeod et al. 2017, which includes Bennett) report empirical or field-specific findings and are used as evidence for particular phenomena, not as the sole justification for the central normative claim; they are thus independent support under the reviewing rules. The skeptic's concern that the persistence of harm under fairness is asserted rather than demonstrated is a legitimate evidentiary limitation, but it concerns the strength of the causal/social argument, not circularity. No equation, uniqueness theorem, ansatz, or renamed empirical pattern is imported from the authors' prior work, so there are no circular steps.
Assumptions & free parameters
assumptions (4)
- domain assumption The operative meaning of fairness in AI ethics is statistical parity or similar-case-same-outcome, and it does not question underlying power structures.
- domain assumption Medicalization and diagnostic gatekeeping are structurally harmful to disabled people, independent of diagnostic accuracy.
- domain assumption Assistive computer vision tools shift judgment from users to technology and privilege vision over other sensory modes, producing harm even when deployed fairly.
- domain assumption The harms cited, such as surveillance misuse, violence against autistic people, and nonvisual sensemaking being undervalued, are representative and causally connected to the technologies discussed.
Cite this review
Pith. "Pith review of What is the Point of Fairness? Disability, AI and The Complexity of Justice." pith.science (2026). https://pith.science/paper/KEEMC32D
@misc{pith2026190801024,
author = {Pith},
title = {Pith review of: What is the Point of Fairness? Disability, AI and The Complexity of Justice},
year = {2026},
howpublished = {\url{https://pith.science/paper/KEEMC32D}},
note = {Machine review of arXiv:1908.01024}
}
read the original abstract
Work integrating conversations around AI and Disability is vital and valued, particularly when done through a lens of fairness. Yet at the same time, analyzing the ethical implications of AI for disabled people solely through the lens of a singular idea of "fairness" risks reinforcing existing power dynamics, either through reinforcing the position of existing medical gatekeepers, or promoting tools and techniques that benefit otherwise-privileged disabled people while harming those who are rendered outliers in multiple ways. In this paper we present two case studies from within computer vision - a subdiscipline of AI focused on training algorithms that can "see" - of technologies putatively intended to help disabled people but, through failures to consider structural injustices in their design, are likely to result in harms not addressed by a "fairness" framing of ethics. Drawing on disability studies and critical data science, we call on researchers into AI ethics and disability to move beyond simplistic notions of fairness, and towards notions of justice.
Reference graph
Works this paper leans on
-
[4]
Sander Begeer, Saloua El Bouk, Wafaa Boussaid, Mark Meerum Terwogt, and Hans M. Koot. 2009. Underdiagnosis and referral bias of autism in ethnic minorities. J Autism Dev Disord. 39, 1 (January 2009), 142-148. DOI: https://doi.org/10.1007/s10803-008-0611-5 5. Hendrik Buimer, Thea Van der Geest, Abdellatif Nemri, Renske Schellens, Richard Van Wezel, and Yan...
arXiv 2010
-
[12]
(2019) Where fairness fails: Data, algorithms, and the limits of antidiscrimination discourse
Anna Lauren Hoffmann. (2019) Where fairness fails: Data, algorithms, and the limits of antidiscrimination discourse. Information, Communication, & Society, 22, 7 (2019), 900-915. 13. Os Keyes. 2018. The Misgendering Machines: Trans/HCI Implications of Automatic Gender Recognition. In Proc. ACM Hum.-Comput. Interact. 2, CSCW, Article 88 (November 2018), 22...
arXiv 2019
-
[34]
World Institute on Disability. 2019. AI and Accessibility (July 2019). Retrieved 23 June, 2019 from https://wid.org/2019/06/12/ai-and-accessibility/ 35. Yuhang Zhao, Cynthia L. Bennett, Hrvoje Benko, Edward Cutrell, Christian Holz, Meredith Ringel Morris, and Mike Sinclair. 2018. Enabling People with Visual Impairments to Navigate Virtual Reality with a H...
arXiv 2019
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.