PaperlayerView exact revision on arXiv

Source-linked AI summary

AI Fairness for People with Disabilities: Point of View

Shari Trewin

arXiv:1811.10670v1cs.AIcs.CY

TL;DR

Disability discrimination in machine-learning systems has received little attention despite its potential for bias. This paper examines disability’s distinctive characteristics and common fairness definitions, arguing that disability fairness differs because disabilities are diverse, dynamic, and not always disclosed.

  • Problem

    Disability discrimination in machine-learning systems has been little explored, despite the potential for bias against people with disabilities.

    Source: p. 1

  • Method

    The paper describes disability’s distinctive characteristics, discusses common fairness definitions, and offers recommendations for AI solutions.

    Source: p. 2

  • Results

    The paper argues that disability fairness differs from fairness for other protected attributes because disability is diverse, dynamic, and not always disclosed.

    Source: p. 3, Measuring Fairness - Group Fairness · p. 5

  • Takeaways & Limitations

    Developers should consider disability implications, involve affected groups where possible, and test performance on edge cases and under-represented populations.

    Source: Measuring Fairness - Individual Fairness · p. 6, Conclusion · p. 7

    Group fairness measures require explicit disability information, can conceal discrimination against subgroups, and cannot accommodate genuine differences between groups.

    Source: Measuring Fairness - Group Fairness · p. 5

Abstract

from arXiv · show

We consider how fair treatment in society for people with disabilities might be impacted by the rise in the use of artificial intelligence, and especially machine learning methods. We argue that fairness for people with disabilities is different to fairness for other protected attributes such as age, gender or race. One major difference is the extreme diversity of ways disabilities manifest, and people adapt. Secondly, disability information is highly sensitive and not always shared, precisely because of the potential for discrimination. Given these differences, we explore definitions of fairness and how well they work in the disability space. Finally, we suggest ways of approaching fairness for people with disabilities in AI applications.

Fairness through unawareness

Avoiding disability data does not ensure fair outcomes, and group- and individual-fairness approaches each face disability-specific limitations; the section urges inclusive testing, bias mitigation, and documentation.

Fairness through unawareness

  • Ignoring disability can protect sensitive information and align with practice, but inaccessible tests can exclude applicants through correlated signals, while absent labels hinder discrimination assessment.For example, long response times on inaccessible tests could systematically exclude applicants using screen readers or magnifiers.Source: Fairness through unawareness · pp. 4–5

Measuring Fairness - Group Fairness

  • Group fairness seeks roughly equal selection rates, but for disability it requires sensitive labels, can hide subgroup harms, is impractical across all groups, and ignores genuine differences.The paper describes this as affirmative action that does not guarantee selection of the most qualified candidates.Source: Measuring Fairness - Group Fairness · p. 5

Measuring Fairness - Individual Fairness

  • Individual fairness gives similar task-relevant applicants similar outcomes without requiring disability labels, but depends on a fair, application-specific similarity metric that disability-affected inputs can bias.Test suites representing a broad range of people with disabilities could probe resulting models.Source: Measuring Fairness - Individual Fairness · p. 5
  • Developers should involve disability groups, test and mitigate bias across the pipeline, scrutinize disability-affected inputs, improve outlier robustness, and document inclusion, tests, mitigations, and limitations.Training data from affected groups can help address concerns, but the paper notes that outlier individuals will remain.Source: Measuring Fairness - Individual Fairness · p. 6

Conclusion

The conclusion calls attention to disability discrimination as an underexamined issue in AI and warns that systems optimized for typical cases may disadvantage disabled people whose experiences appear as edge cases. It argues that people with disabilities should be involved in developing systems that affect human lives, including to help anticipate implications and test performance.

  • Disability discrimination in AI has received little research attention compared with gender, racial, and age bias.Source: Conclusion · p. 6
  • Pure ML optimization for typical cases can disadvantage disabled people, whose diverse experiences may appear as data edge cases, as Treviranus (Treviranus, 2017) cautions.Source: Conclusion · p. 6
  • For systems affecting human lives, broad stakeholder involvement, including people with disabilities, can help developers anticipate implications and test edge-case and under-represented-population performance.Source: Conclusion · pp. 6–7
Loading 1811.10670v1…