Fairness of AI for People with Disabilities: Problem Analysis and Interdisciplinary Collaboration
Jason J. G. White · 2020 · SIGACCESS Accessibility and Computing · doi:10.1145/3386296.3386299
Summary
This paper provides a philosophical analysis of the fairness challenges that machine learning-based AI poses for people with disabilities, arguing that these challenges demand unprecedented interdisciplinary collaboration across applied ethics, human rights law, disability studies, privacy, and AI research. White identifies several interconnected problems. First, the underrepresentation problem: people with disabilities are systematically absent from AI training datasets, leading to poor performance on tasks like speech recognition for people with speech-related disabilities or image recognition for users with atypical interaction patterns. Collecting disability-specific data to remedy this raises serious privacy concerns, since the data needed to improve AI systems could also be used to identify disability status and enable discrimination. A second form of underrepresentation occurs when people with disabilities are excluded from the design, development, and evaluation of AI systems. The paper also examines how automated decision-making systems trained on biased historical data can compound existing societal injustices — for example, an algorithm screening job applications might penalize employment gaps attributable to disability, reinforcing rather than correcting discrimination. White draws on philosophical frameworks from Nussbaum's capabilities approach and Rawls' theory of justice to analyze these issues, noting that disability creates fundamental challenges for standard fairness metrics because it affects people's capacity to convert resources into well-being in ways that cannot be reduced to simple numerical measures.
Key findings
White identifies four core fairness challenges. The training data underrepresentation problem parallels challenges faced by linguistic minorities but is complicated by the privacy risks of disability-related data collection — people must disclose disability status to be included, creating vulnerability to discrimination. The counterfactual analysis framework commonly used to evaluate algorithmic fairness (asking whether a decision would have been different absent the protected characteristic) is especially difficult to apply to disability because of the enormous heterogeneity of disabled people's experiences and circumstances. White raises the critical question of whose problems AI is solving: while AI advances in computer vision and NLP hold enormous potential for people with disabilities, there is a risk that research investment flows primarily toward profitable mainstream applications rather than disability-specific needs, constituting a form of injustice through resource allocation. The paper argues that privacy-preserving techniques like differential privacy, federated learning, and homomorphic encryption could help resolve the tension between data collection and privacy protection. Throughout, White emphasizes that the moral question of which decisions should be automated at all must remain central to the discussion.
Relevance
This paper is essential reading for anyone working at the intersection of AI and accessibility. It provides a rigorous ethical framework for thinking about AI fairness that goes beyond the typical race-and-gender focus of algorithmic bias research to address disability specifically. For organizations deploying AI systems, the paper raises practical questions: Are people with disabilities represented in training data? Are they involved in design and evaluation? Could automated decisions compound disability-based disadvantage? For accessibility practitioners, the underrepresentation problem has direct implications — AI-powered assistive technologies like speech recognition, image description, and predictive text may perform poorly for the very users who need them most. The call for interdisciplinary collaboration between technologists, ethicists, legal scholars, and disability advocates offers a roadmap for more equitable AI development. White's emphasis that disability fairness shares commonalities with — but is distinct from — other forms of algorithmic bias helps position disability within broader social justice conversations about AI governance.
Tags: AI fairness · algorithmic bias · disability · social justice · ethics · machine learning · policy · underrepresentation