Jump to Content

The Deskilling of Domain Expertise in AI Development

Nithya Sambasivan
Rajesh Veeraraghavan
SIGCHI, ACM (2022) (to appear)
Google Scholar


Domain experts, like farmers and radiologists, play a crucial role in dataset collection for models, especially in low-resource settings. However, we know little about the labour process and how domain experts are leveraged in dataset and model development. Based on 68 interviews with AI developers building for low-resource contexts, we find that developers typecast domain experts as data collectors, exacting repeatable and standardized tasks to create datasets. Developers attributed poor data quality to poor work practices of field experts, through conceptions of: domain expert as corrupt, lazy, non-compliant, and dataset itself. Developers resorted to disciplinary interventions to manage experts, in the hope of improving dataset quality. We argue that AI development in low-resource areas \textit{deskills expertise} of local domain experts in service of machine intelligence. We argue for a shift to enrolling domain experts as full-range technical experts, co-creating datasets and models, and accounting for surplus labour in AI models.