Persons with disabilities are statistical minorities by definition, AI is generally less accurate and fair towards them because it is optimized for the average or typical.
Disability-related language is often misclassified as harmful or negative, leading AI systems to stigmatize disability and suppress related content.
People with disabilities are often underrepresented in training data, so AI systems are less likely to reflect their experiences or meet their needs.
AI systems trained on typical behaviour and bodies may incorrectly flag disability-related patterns as suspicious or fraudulent.
AI systems learn from past data and may continue patterns of exclusion by filtering out people with disabilities, even when disability is not explicitly stated.
Repeated small AI decisions can build up over time, leading to significant harm for people with disabilities across multiple points of interaction.
Tools used to build and review AI are often inaccessible, limiting the ability of people with disabilities to participate in using and improving these systems.
AI systems may generate incorrect information with high confidence, creating risks for people who rely on these outputs and may not be able to verify them.
People with disabilities may face a trade-off between sharing information to access support and risking misuse, discrimination, or re-identification.
Heavy reliance on AI can lead to standardized outputs that replace a person’s own voice and reduce the diversity of lived experiences represented.
AI tools may work less well for some people with disabilities than others, even when they are designed to remove barriers.
Assistive AI tools may be expensive or restricted, leaving people with disabilities vulnerable to price changes or loss of access.
Low-paid AI data work may exploit people with disabilities in ways that reinforce economic inequality and risk worsening health conditions.
Harm against persons with disabilities is determined to be anecdotal or statistically insignificant when using statistical determination of negative impact or risk. Mitigation strategies that rely on statistical reasoning are subject to bias against statistical outliers.