Summary
This technical report discusses the risks associated with the use of synthetic data in machine learning, particularly focusing on facial recognition technology. It identifies two primary risks: the potential for false confidence in dataset diversity and the circumvention of consent for data usage. The report outlines how synthetic data can create an illusion of diversity in datasets, which may not address the underlying cultural and qualitative factors that contribute to representation in real-world data. Additionally, it examines the implications of using synthetic data to bypass consent regulations, highlighting the importance of consent as mandated by the U.S. Federal Trade Commission. The authors emphasize that while synthetic data can augment datasets, it also complicates ethical practices and governance in data collection and model development. The report aims to contribute to ongoing discussions about responsible synthetic data usage and its impact on algorithmically-mediated harm.
