ABSTRACT
Our collaboration seeks to demonstrate shared interrogation by exploring the ethics of machine learning benchmarks from a socio-technical management perspective with insight from public health and ethnic studies. Benchmarks, such as ImageNet, are annotated open data sets for training algorithms. The COVID-19 pandemic reinforced the practical need for ethical information infrastructures to analyze digital and social media, especially related to medicine and race. Social media analysis that obscures Black teen mental health and ignores anti-Asian hate fails as information infrastructure. Despite inadequately handling non-dominant voices, machine learning benchmarks are the basis for analysis in operational systems. Turning to the management literature, we interrogate cross-cutting problems of benchmarks through the lens of coupling, or mutual interdependence between people, technologies, and environments. Uncoupling inequality from machine learning benchmarks may require conceptualizing the social dependencies that build structural barriers to inclusion. © 2022 IEEE Computer Society. All rights reserved.