Developing Pro-Social AI Training Datasets Through Data Workers' Critical Perspectives

The massive datasets used to train AI models frequently contain offensive and harmful entries that are only caught during the system’s later performance. The data workers who curate such datasets are experts in datasets’ contents, but silenced by horrible labor conditions and lack of respect by their employers. In my dissertation, I study how to build 1) workplaces, 2) workflows, and 3) tools to elicit and embrace data workers’ observations in dataset development. My goals are creating datasets safe to use to train AI systems and developing a more pro-social data labor paradigm.