Including Disability in Datasets and AI Development

Abstract

Advances in artificial intelligence (AI) promise to remove accessibility barriers experienced by disabled people. Data is often a prerequisite for this promise. However, datasets sourced from disabled people (i.e., accessibility datasets) are scarce, limiting innovation and raising concerns of harm for this community. Inclusivity issues, including lack of data representativeness, tend to rise in the development. Too often, the teams defining objectives and model requirements for developing AI-infused solutions lack members with lived experience of disability. More so, existing practices to include disabled people in the AI development lifecycle often come with ethical, practical, and methodological dilemmas. The overarching goal of my research is to realize this promise of AI for accessibility. Through a weave of threads and a mixed-methods approach, this dissertation proposal aims to mitigate issues of inclusivity by ascertaining the gaps and addressing the misalignments of datasets and AI development practices with disabled people’s needs and contexts. Thread I. Underlining the importance of data, I discover and question the norms related to including disability in datasets. I conduct a meta-analysis of accessibility datasets over the last four decades to ascertain the gaps in data collection and sharing practices and their result in intersectional representation. Thread II. To address misalignment of existing norms with the disability community, I surface people’s perspectives on data collection and sharing for increasing disability representation in AI datasets. With a focus on the blind community, I conduct contextual interviews, followed by a co-design study to foster active participation of blind people to shape future data stewardship practices. Thread III. Moving beyond datasets, I expand my research to broader AI development including problem formulation and evaluation. With a focus on the Deaf community, I conduct surveys to ascertain the gaps in expectations between machine learning practitioners and Deaf/deaf people in AI for sign language. I also address misalignment via paired co-design studies to foster collaboration involving the Deaf community in AI teams.