Browse by Benefits

Risk:

AI systems are optimized for the average user, structurally excluding persons with disabilities who, by virtue of disability, exist at the statistical edges and vary substantially from one another. As a result, AI tools tend to perform worse for the individuals who depend on them the most. The risks and harms experienced by persons with disabilities are then frequently dismissed as statistically insignificant or merely anecdotal due to the smaller population size and the uniqueness of each individual case.

Mitigation:

To address this, disability-disaggregated impact assessments should be required before deployment, and AI tools should be tested against statistical edges as opposed to averages alone. Reporting worst-case accuracy for subgroups alongside average accuracy helps to ensure that exclusion at the edges is surfaced, as opposed to hidden.

Risk:

AI text and image classifiers are often trained on datasets which perpetuate social biases against persons with disabilities; as a result, disability-related terms tend to be flagged as toxic, negative, or harmful. This produces stigmatizing, patronizing, or pity-driven representations of disability, as well as content moderation systems that suppress disability-related speech.

Mitigation:

Toxicity classifiers should be audited before deployment to remove disability-language scoring, and disability sentiment benchmarks should be incorporated into model evaluation. Partnering with disabled creators and disabled persons' organizations (DPOs) is essential to ensure that representation testing reflects lived experiences.

Risk:

Training datasets rarely represent persons with disabilities adequately; as a result, AI outputs tend to default to a non-disabled average (even when explicitly prompted to depict disability) and image generators frequently omit assistive devices and disability-relevant context. This issue is particularly detrimental for low- and middle-income countries (LMICs), where most large language models, text-to-speech systems, and assistive AI tools are trained on a handful of major languages. Resultingly, this excludes users of under-resourced sign languages (i.e., Mexican Sign Language, Jamaican Sign Language, etc.) and minority spoken languages from accessing functional tools in their own language.

Mitigation:

Disability representation benchmarks should be required for generative AI, and minority-language and sign-language coverage should be a condition of procurement. Funding LMIC-led model development for under-resourced languages (particularly sign-language AI in regions where interpreter shortages are most severe) is essential in addressing this gap.