A Framework for Ethical Data Removal from Language Resources: An Example from Low-Resource Language Communities

Abstract

As data resources such as Common Crawl, Mozilla Common Voice, The Pile, and LAION increasingly serve as the raw material for foundational models, the ethical implications of data collection practices become more complex. This paper addresses some of the growing concerns regarding data removal from said resources. Further, this paper presents a framework for data resource hosts to follow when deciding when to remove data. It also presents a process for individuals and/or communities to follow when seeking to have their data removed. Finally, numerous technical challenges and societal trade-offs are addressed.