Leveraging Digital Accessibility Using Generative AI

Abstract

Generative AI (GenAI), and especially large language models (LLMs), offer intriguing opportunities to improve accessibility in the digital world. In this position paper, we present our research findings and ideas for next steps in two areas of ongoing work in this field. First, we show that LLMs can adjust HTML to mitigate deceptive patterns, manipulative designs in online user interfaces, without prior training. For this, we prompted GPT-4o with HTML code, asked it iteratively to make it less manipulative, and manually analyzed 2,600 redesigns that demonstrate the potential of this approach as a technical countermeasure against deceptive patterns. Second, we outline how GenAI approaches can augment screen readers used by blind and low-vision users by enabling natural, conversational interactions and generating missing metadata. A particular advantage of GenAI for this use case lies in its capability to adapt to individual user preferences and needs.