How an LLM can improve automatic web accessibility validation?

Digital accessibility is important since it allows all people, including those with disabilities, to interact and access the desired information available on the Web. The W3C WCAG guidelines provide a rich set of indications about how to obtain it. Over time, they have become very extensive in order to consider the many possible cases. Manual checking of all the corresponding techniques is impossible; thus, interest in the support provided by automatic tools is increasing. However, current validation tools sometimes have several limitations in their analysis, which still require considerable manual intervention to validate several accessibility techniques. Large Language Models (LLMs) present an opportunity to address such cases. In this paper, we report on an investigation that focused on exploiting the functionality made available by the GPT 4o APIs to address such cases. We report on the types of prompting techniques used for this purpose, how they have been exploited, for which accessibility techniques, and how they have been validated. The results provide useful indications for understanding the role of large language models for accessibility validation.