Toward Responsible Disability Representation and Accessibility AI Applications

Dr. Cynthia Bennett

Abstract:

Enthusiasm to apply AI to solve real-world challenges including remediating inaccessible media is increasing more than ever before with the rapid development and popularization of generative AI products. They may scale tomes of inaccessible content and offer accessible outlets for creative expression. However, the generative models underlying these tools and products exhibit bias, disability misrepresentation being among the most under-studied. Further, successful accessibility provisions are nuanced to individuals' needs and accessibility is an interdependent process, impacted by the environment, people involved, and other contextual factors. In short, accessibility is human;  many accessibility provisions require humans to spend time and to learn specific skills. This raises questions about how the quality and responsibility of accessibility will be impacted by automating its development at scale with generative AI solutions.

In this talk I will discuss two projects at the intersection of responsible AI, people with disabilities, and accessibility. Through the first project, I will share evaluations of disability representation in AI images, by people who have disabilities. Participants prompted text-to-image AI tools in their own language for disability, health, and accessibility-related images they wanted to see, and they could iterate on prompts when output images did not meet their expectations. I will share categories of their language used to specify disability, health, and accessibility, intersectional representational harms participants identified in output images, and discuss greater considerations for generative AI evaluations rooted in community expertise.

Through the second project , I will share interviews we conducted with audio description (AD) professionals, who write, narrate, quality control, and produce narration tracks mixed into videos and live theater which describe important visual events, to make the media accessible to blind and low vision (BLV) audiences. For decades, AD has been standardized and developed by people. Yet many videos are still produced without AD tracks, and generative vision-language models bring promise that AD could be created at scale. However, AD professionals and users are concerned about the quality tradeoffs consumers are required to accept, and their accessibility ethics include impacts on labor and the appropriate for certain AD to be automated. To this end, Advocates assert that quality AD is also culturally competent and blind-centered. The former refers to AD which accurately and respectfully describes cultural aspects of the media in a way appropriate to the intended audience, and blind-centered AD refers to significant involvement by BLV professionals. These values clash with AI models which exhibit white and western biases, among others, and foregrounding BLV experts becomes challenging as more aspects of AD production are automated. I will share canonical AD workflows from our participants which exhibit expertise they bring to its development, and their needs for technical intervention that supports rather than erases their expertise. 

Bio: Dr. Cynthia Bennett is a senior research scientist at Google Research. Her research is about making technology-mediated experiences, such as those leveraging generative AI, accessible to and representative of people with disabilities while mitigating harmful applications. Prior, Bennett was a researcher at Apple and a postdoctoral Research Fellow at Carnegie Mellon University, after receiving her Ph.D. in Human Centered Design and Engineering from the University of Washington. Bennett's research has been recognized with awards from top professional organizations, including the 2025 Skip Ellis Early Career Award from the Computing Research Association. She is also a disabled woman scholar committed to raising participation of people with disabilities in the tech industry.

Zoom Link: https://cuboulder.zoom.us/j/95307227064