Purpose/objective: Text-to-image (TTI) systems are artificial intelligence (AI) models that incorporate large amounts of data to produce high-resolution images. Although research has documented racial/ethnic and gender bias in TTI, little has examined disability bias. This study compared generated images of disabled people with no prompted setting to images of disabled individuals in health care settings.
Research method/design: OpenAI's DALL-E-3 TTI generated 50 images for each of the following prompts: (a) "person with a disability," (b) "patient with a disability," (c) "doctor with a disability," and (d) "doctor with a disability and a patient without a disability." We calculated DALL-E's success in generating prompted images and coded disability type and demographics.
Results: When prompted to create a "person with a disability," DALL-E-3 was 100% successful, with a wide diversity of disabilities. When prompted to create a "patient with a disability," DALL-E-3 was similarly 100% successful, although 70% of images portrayed an individual with a stereotypical physical disability. When prompted to create a "doctor with a disability," DALL-E-3 did with 92% accuracy: 94% had a physical disability and 6% a sensory disability; no other disability types were portrayed. When prompted to create a "doctor with a disability and a patient without a disability," in 64% of cases, DALL-E-3 generated images of doctors without disabilities, and 70% portrayed a disabled patient instead.
Conclusions/implications: Disability diversity decreases dramatically when AI-generated images place disabled people in a medical environment. As TTI generation grows more ubiquitous, further work by model developers to mitigate representational harms is vital. (PsycInfo Database Record (c) 2025 APA, all rights reserved).
扫码关注我们
求助内容:
应助结果提醒方式:
