Disability Expertise in Making AI Work

• In Event: AI and Feminist STS - II

Abstract

People with disabilities are typically imagined as victims, beneficiaries, or inspirations of AI. This paper examines an under-discussed role that disabled people play in AI systems – labor. I focus on a China-based team of disabled data workers who sort, label, and categorize training data for voice-activated AI (VAI) systems. Based on remote interviews and preliminary pre-pandemic fieldwork, I show how these blind, low-vision, or physically impaired workers, through centering the ethics of care, access, and interdependence, reshape work in AI production.

I argue that AI developers in China profit from systemic ableism and disability expertise, not the other way around. The inherent ambiguity of human intention categorization that current VAI systems rely on, mobilizes a stable, trained workforce of data taggers, which disabled people in the absence of better structural opportunities end up fulfilling. Yet not all “disability + tagging” programs succeed.

In this study, a disability NGO organized the workers to negotiate and improve the working conditions of data annotation. Specifically, they deploy situated knowledge of “crip time” management, and the co-creation of access. These forms of “disability expertise,” grounded in the heterogenous lived experience of disability, I argue, are key to improving the work performance and work experience of AI data processing.

Building upon scholarship of feminist STS and crip technoscience, this paper calls for attention to the knowledge and expertise of disabled people in human-AI systems. Centering the co-creation of enabling conditions rather than essentialized body-minds, disability expertise offers a new methodology for imagining counterhegemonic AI.

Author

• Di Wu, Massachusetts Institute of Technology (MIT)