An approach to increasing nonvisual access to information for people who are blind or low vision is artificial intelligence (AI), which promises automation and scalability. However, research and media reports continue to illuminate AI-bias and malicious applications. As these harms tend to impact people who already experience marginalization, the ethics of applying AI to solve perennial accessibility challenges is complicated. This impact is compounded by the conversational communication style of recent generative AI advancements, which may make false information read as convincing to audiences who experience accessibility barriers to confirming ground truth (e.g., people who are blind or low vision may not be able to decide whether an AI-generated image description is accurate since the description remediates fundamental access).

In this talk I will overview one project I led concerning one facet of information accessibility—how people are represented in human and AI-generated alternative (alt) text descriptions. I will share findings from interviews I conducted with blind people who rely on alt text to understand visual information and who also identified as a minoritized race or gender shown to be disproportionately misrepresented and harmed by algorithmic decision-making. From their experiences and perspectives, triangulated with recent disability activism and artistic visual description movements, I synthesized human representation alt text design considerations. I will conclude the talk with brief windows into my current research and collaborations which explores how context might be represented in datasets, towards richer and more situated alt text.