Study finds disability stereotypes in AI chatbots

A peer-reviewed study by Special Olympics with Oregon State University found that five leading AI chat systems consistently produced stories portraying people with intellectual disabilities as more dependent, childlike and in need of supervision than people without disabilities. Researchers warned that such patterns could contribute to automated ableism in systems used for work, healthcare, education and access to information.

The study, published in Disability and Health Journal, tested commonly used Large Language Models including GPT-4-Turbo by asking them to generate 25,000 short stories about everyday situations. Some prompts identified a person as having ID and others did not, while GPT-4-Turbo was used to analyze recurring patterns such as paternalism, dependence, infantilization, inspirational framing, age representation and negative perception.

Findings were consistent across all five models. Stories about people with ID were in some cases over 100 times more likely to use a paternalistic tone, often framed people with ID as needing help with decisions, and described them as younger than peers without ID. Special Olympics noted that it serves athletes aged 8 years old to 80 years old, underscoring concerns about narrow portrayals.

Special Olympics said AI can also improve accessibility by simplifying medical records, creating text-to-image communications, providing way-finding directions, guiding social skill acquisition and supporting decision-making. The organization called on developers, researchers and institutions to include people with intellectual and developmental disabilities in AI design and bias mitigation work.